Awesome Unitree Robots 中文

📝 Introduction

A curated list of high-quality open-source projects, research implementations, and developer tools for Unitree Robotics platforms—including Go2, G1, H1, H1-2, B2, and Aliengo. This collection empowers robotics researchers and engineers with reproducible codebases, simulation environments, and real-world deployment examples spanning locomotion, teleoperation, SLAM, and reinforcement learning. Emphasis is placed on projects with public code, clear hardware integration, and active community or academic support. Key terms: Unitree-Go2, Unitree-G1, Unitree-H1, Unitree-H1-2, Unitree-B2, Unitree-Aliengo.

🎯 Scope

• Official SDKs, simulators, and firmware from Unitree
• Academic papers with open-source implementations targeting Unitree robots
• Community-driven tools for control, perception, and autonomy
• Projects validated in simulation (Isaac Lab, MuJoCo, Isaac Gym) and/or on real hardware

✅ Inclusion Criteria

• Must include publicly accessible code (GitHub/GitLab) or peer-reviewed paper (ArXiv/conference)
• Explicit support or adaptation for at least one Unitree model (e.g., Go2, H1, G1)
• Demonstrates technical novelty or engineering utility (e.g., Sim2Real transfer, whole-body MPC, low-latency teleop)
• Published or updated within the last 3 years (2022–2025)

👥 Audience

Robotics developers, graduate researchers, and autonomy engineers building on Unitree’s quadruped and humanoid platforms.

📚 Table of Contents

• Introduction- Scope- Inclusion Criteria- Audience- Papers & Research
• Projects
• Related Awesome Lists

🔥 Papers & Research

• "HUSKY: Humanoid Skateboarding System via Physics-Aware Whole-Body Control" (2026) PDF — The paper introduces HUSKY, a learning-based framework for humanoid skateboarding that combines physics-aware whole-body control with Adversarial Motion Priors (AMP) to enable dynamic maneuvering on an underactuated skateboard. It explicitly models the coupling between board tilt and truck steering, and implements lean-to-steer and pushing behaviors on the Unitree G1 humanoid robot. Real-world experiments demonstrate stable and agile skateboarding, showcasing the system’s ability to handle hybrid contact dynamics and non-holonomic constraints, with significant implications for dynamic human-robot interaction tasks.

• "Embodiment-Aware Generalist Specialist Distillation for Unified Humanoid Whole-Body Control" (2026) PDF — The paper introduces EAGLE, an iterative generalist-specialist distillation framework for unified whole-body control across diverse humanoid robots without per-robot reward tuning. It explicitly validates the method on real-world Unitree H1 and G1 platforms, alongside simulation experiments on five robots, demonstrating high tracking accuracy and robustness. This work directly contributes to scalable, fleet-level control of Unitree humanoids and provides a practical pathway for deploying a single policy across heterogeneous hardware.

• "Training and Simulation of Quadrupedal Robot in Adaptive Stair Climbing for Indoor Firefighting: An End-to-End Reinforcement Learning Approach" (2026) PDF — This paper presents a two-stage end-to-end reinforcement learning framework for enabling the Unitree Go2 quadruped robot to perform adaptive stair climbing in complex indoor environments relevant to firefighting scenarios. The approach first trains the robot in Isaac Lab using an abstract pyramid-stair terrain and then transfers the policy to realistic staircases (straight, L-shaped, spiral) within the same simulator. By integrating navigation and locomotion through a centerline-based formulation without hierarchical planning, the method demonstrates strong generalization across diverse stair geometries using only local perception.

• "PRISM: Performer RS-IMLE for Single-pass Multisensory Imitation Learning" (2026) PDF — PRISM introduces a single-pass multisensory imitation learning policy using a Performer-based architecture and a rejection-sampling variant of IMLE to enable real-time, high-frequency control across diverse sensing modalities. The method is validated on real-world hardware including the Unitree Go2 quadruped equipped with a 7-DoF D1 arm, demonstrating superior success rates (10–25% higher) over diffusion policies in loco-manipulation tasks like pre-manipulation parking and high-precision insertion while maintaining 30–50 Hz control. PRISM also shows strong performance in large-scale simulated benchmarks, offering both improved task success and smoother trajectories.

• "RAPT: Model-Predictive Out-of-Distribution Detection and Failure Diagnosis for Sim-to-Real Humanoid Robots" (2026) PDF — The paper introduces RAPT, a lightweight, self-supervised monitoring system for real-time out-of-distribution (OOD) detection and failure diagnosis in sim-to-real humanoid robot deployment. It explicitly evaluates the Unitree G1 humanoid across four complex tasks, using a probabilistic spatio-temporal manifold learned from simulation to detect anomalies at 50Hz and provide interpretable, per-dimension deviation signals. RAPT also features an automated root-cause analysis pipeline combining gradient-based saliency and LLM reasoning to generate semantic failure explanations, offering practical value for safe and diagnosable real-world humanoid operation.

• "HumanX: Toward Agile and Generalizable Humanoid Interaction Skills from Human Videos" (2026) PDF — HumanX introduces a full-stack framework that enables humanoid robots to learn agile, generalizable interaction skills directly from human videos without task-specific rewards. The method combines XGen for generating physically plausible robot interaction data from video and XMimic for unified imitation learning. It successfully transfers 10 diverse skills—such as basketball jumpshots and sustained passing—to the physical Unitree G1 humanoid with zero-shot sim-to-real transfer, achieving over 8× higher generalization success than prior work.

• "Toward Reliable Sim-to-Real Predictability for MoE-based Robust Quadrupedal Locomotion" (2026) PDF — This paper introduces a Mixture-of-Experts (MoE) reinforcement learning framework for robust quadrupedal locomotion and RoboGauge, a sim-to-real predictability suite that evaluates policy transferability using proprioceptive signals. The approach is validated on the Unitree Go2, demonstrating successful deployment on diverse unseen terrains like snow, sand, stairs, and 30 cm obstacles, as well as high-speed locomotion up to 4 m/s with an emergent stable gait. The work directly uses the Unitree Go2 as its experimental platform, making it highly relevant to the Awesome Unitree Robots list.

• "ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control" (2026) PDF — ZEST (Zero-shot Embodied Skill Transfer) is a reinforcement learning framework that enables zero-shot transfer of dynamic, whole-body skills to humanoid robots using diverse motion sources like mocap, video, and animation. The method is validated on multiple platforms including Boston Dynamics' Atlas and Unitree G1, demonstrating agile behaviors such as dancing and box-climbing without platform-specific tuning. By leveraging simulation-only training with adaptive sampling and an assistive wrench curriculum, ZEST achieves robust Sim2Real transfer, offering a scalable approach for athletic robot control.

• "kevinkawchak/clinical-trial-rl-unitree-g1: v0.4" (2026) — Zenodo (CERN European Organization for Nuclear Research) — This repository presents reinforcement learning (RL) trained policies for the Unitree G1 humanoid robot, achieving improved locomotion capabilities over prior versions. It includes 1500-iteration RL weights, an extended Jupyter notebook demonstrating three key behaviors—careful navigation with obstacle avoidance, patient approach with pausing, and dynamic movement adaptation—and accompanying videos. The work directly uses the Unitree-G1 as the experimental platform, focusing on clinically inspired mobility tasks, and provides practical tools for sim-to-real transfer in healthcare robotics scenarios.

• "PILOT: A Perceptive Integrated Low-level Controller for Loco-manipulation over Unstructured Scenes" (2026) PDF — arXiv (Cornell University) — The paper introduces PILOT, a unified reinforcement learning framework for perceptive loco-manipulation that integrates terrain-aware locomotion and whole-body manipulation in unstructured environments. It employs a cross-modal context encoder and a Mixture-of-Experts policy architecture to enhance foot placement accuracy and motion specialization. The method is validated both in simulation and on the physical Unitree G1 humanoid robot, demonstrating improved stability, command tracking, and traversability, making it a strong candidate for real-world deployment on Unitree platforms.

• "kevinkawchak/clinical-trial-rl-unitree-g1: v0.2" (2026) — Zenodo (CERN European Organization for Nuclear Research) — This repository presents reinforcement learning (RL) policies trained for the Unitree G1 humanoid robot, focusing on clinical trial scenarios such as careful navigation, patient approach, and equipment transport. It includes three RL models trained over 3000 iterations with improved mean reward and reduced fall rates, along with Jupyter notebooks and model checkpoints. The work directly uses the Unitree-G1 as the experimental platform, offering practical Sim2Real-ready policies for healthcare robotics applications.

• "kevinkawchak/clinical-trial-rl-unitree-g1: v0.3" (2026) — Zenodo (CERN European Organization for Nuclear Research) — This repository presents reinforcement learning (RL) experiments focused on training the Unitree G1 humanoid robot to perform clinical tasks such as patient approach and careful navigation. Using RL, the project demonstrates significant improvements in whole-body stability by 3000 iterations, achieving stable walking with episode lengths of ~1000 steps and a peak reward of ~45.80. The work highlights task-specific behaviors like subtle movement control but notes limitations in equipment transport utility.

• "HumanoidVLM: Vision-Language-Guided Impedance Control for Contact-Rich Humanoid Manipulation" (2026) PDF — arXiv (Cornell University) — This paper introduces HumanoidVLM, a vision-language-guided framework that enables the Unitree G1 humanoid robot to adapt its impedance control and gripper configuration based on egocentric visual input. By combining a vision-language model with a FAISS-based Retrieval-Augmented Generation (RAG) module, the system retrieves task-appropriate stiffness-damping parameters and grasp angles from custom databases, executing them via a task-space impedance controller. Real-world experiments on the Unitree G1 demonstrate 93% retrieval accuracy and stable compliant manipulation across 14 scenarios, showing a practical path toward semantic, adaptive humanoid control.

• "Vision-Language Models on the Edge for Real-Time Robotic Perception" (2026) PDF — arXiv (Cornell University) — This paper investigates edge deployment of Vision-Language Models (VLMs) for real-time robotic perception using the Unitree G1 humanoid as an embodied testbed. The authors implement a WebRTC-based pipeline to stream multimodal data to an Open RAN/MEC edge node, comparing cloud versus edge inference for LLaMA-3.2-11B-Vision-Instruct and evaluating the compact Qwen2-VL-2B-Instruct model. Their results demonstrate that edge deployment reduces latency while maintaining competitive accuracy, offering practical insights for deploying VLMs on resource-constrained humanoid platforms like the Unitree G1.

• "FocusNav: Spatial Selective Attention with Waypoint Guidance for Humanoid Local Navigation" (2026) PDF — arXiv (Cornell University) — The paper introduces FocusNav, a spatial selective attention framework for humanoid local navigation that combines waypoint-guided perception with stability-aware decision-making. It is explicitly evaluated on the Unitree G1 humanoid robot, demonstrating improved navigation success, collision avoidance, and motion stability in dynamic and complex environments. The method’s core components—Waypoint-Guided Spatial Cross-Attention and Stability-Aware Selective Gating—are validated through real-world experiments on Unitree G1, offering practical value for robust humanoid navigation.

• "FRoM-W1: Towards General Humanoid Whole-Body Control with Language Instructions" (2026) PDF — arXiv (Cornell University) — FRoM-W1 is an open-source framework for language-driven whole-body control of humanoid robots, featuring H-GPT for natural language-conditioned human motion generation and H-ACT for robot-specific motion retargeting and reinforcement learning-based control. The system is explicitly evaluated on Unitree H1 and G1 robots, demonstrating high-fidelity Sim2Real transfer through a modular deployment pipeline. This work provides a generalizable, instruction-following control architecture with strong practical relevance for Unitree humanoid platforms.

• "Learning Robot Locomotion from Diverse Datasets" (2026) — TUprints — This paper proposes a GPT-style generative model that learns quadrupedal locomotion by retargeting diverse motion datasets—including dogs, horses, and other robots—onto the Unitree Go2 and A1 platforms. The method tokenizes motion sequences and conditions generation on gait type and duration, enabling the synthesis of natural, varied gaits. Using a low-level policy, the approach is validated in simulation on the Unitree Go2, demonstrating realistic and controllable locomotion behaviors with potential for Sim2Real transfer.

• "M-SEVIQ: A Multi-band Stereo Event Visual-Inertial Quadruped-based Dataset for Perception under Rapid Motion and Challenging Illumination" (2026) PDF — arXiv (Cornell University) — The paper introduces M-SEVIQ, a novel multi-band stereo event visual-inertial dataset specifically designed for agile quadrupedal robot perception under rapid motion and extreme lighting. It was collected using a Unitree Go2 robot equipped with synchronized stereo event cameras, a frame-based camera, IMU, and joint encoders, providing over 30 real-world sequences across varying speeds and illumination conditions. The dataset includes precise calibration data for sensor fusion and supports research in robust perception, semantic segmentation, and multi-modal vision on legged robots.

• "World-Coordinate Human Motion Retargeting via SAM 3D Body" (2025) PDF — arXiv (Cornell University) — This paper presents a lightweight framework for recovering world-coordinate human motion from monocular videos and retargeting it to the Unitree G1 humanoid robot. It uses SAM 3D Body as a frozen perception module and Momentum HumanRig (MHR) as an intermediate representation, enforcing temporal consistency, smoothing via sliding-window optimization, and recovering physically plausible global trajectories with a contact-aware model. The reconstructed motion is retargeted to the Unitree G1 using a kinematics-aware two-stage inverse kinematics pipeline, demonstrating stable trajectories and reliable robot motion generation from monocular input.

• "Semantic Co-Speech Gesture Synthesis and Real-Time Control for Humanoid Robots" (2025) PDF — arXiv (Cornell University) — This paper presents an end-to-end framework for generating semantically meaningful co-speech gestures from speech and executing them in real time on the Unitree G1 humanoid robot. It combines a semantics-aware gesture synthesis module—using LLMs and Motion-GPT—with a high-fidelity imitation learning controller (MotionTracker) and General Motion Retargeting (GMR) to bridge human-to-robot embodiment gaps. The system enables the Unitree G1 to perform expressive, balanced, and rhythmically coherent gestures synchronized with speech, offering a complete pipeline for real-world deployment of natural non-verbal communication in humanoids.

• "E-SDS: Environment-aware See it, Do it, Sorted - Automated Environment-Aware Reinforcement Learning for Humanoid Locomotion" (2025) PDF — arXiv (Cornell University) — E-SDS introduces an environment-aware reinforcement learning framework that combines vision-language models with real-time terrain sensing to automatically generate reward functions for humanoid locomotion. The method is explicitly evaluated on the Unitree G1 robot across four challenging terrains, demonstrating superior performance—most notably enabling successful stair descent where manual and non-perceptive baselines failed—and reducing velocity tracking error by over 50%. By cutting reward design time from days to under two hours, E-SDS significantly enhances the efficiency and robustness of policy training for real-world Unitree humanoid deployment.

• "Humanoid Localization and Fault Tolerant Control via Nested Control Barrier Functions" (2025) — Open Scholarship Institutional Repository (Washington University in St. Louis) — This paper addresses localization drift and sensor faults in the Unitree G1 humanoid robot by developing a drift-free Extended Kalman Filter (EKF) that fuses IMU, joint encoders, and internal odometry within a center-of-mass motion model. Building on this estimator, the authors propose a nested Control Barrier Function (CBF) architecture with a bank of observers to detect sensor faults and dynamically adjust safety constraints via a quadratic program. The approach ensures collision avoidance and formal safety guarantees even under persistent sensor biases, directly enhancing the reliability of CLF–CBF controllers on the Unitree G1 platform.

• "Beyond Model Jailbreak: Systematic Dissection of the "Ten DeadlySins" in Embodied Intelligence" (2025) PDF — arXiv (Cornell University) — This paper presents the first comprehensive security analysis of the Unitree Go2 platform, identifying ten critical cross-layer vulnerabilities—dubbed the 'Ten Sins of Embodied AI Security'—spanning wireless, core modules, and external interfaces. Through techniques like BLE sniffing, APK reverse engineering, and cloud API testing, the authors expose systemic flaws such as hardcoded keys, missing TLS validation, and insecure firmware access that enable full device hijacking. The work underscores that securing embodied AI demands holistic system-level hardening beyond language model alignment, offering actionable guidelines for robust robot platform design.

• "From Generated Human Videos to Physically Plausible Robot Trajectories" (2025) PDF — arXiv (Cornell University) — The paper introduces GenMimic, a physics-aware reinforcement learning policy that enables humanoid robots to imitate human actions from noisy, generated videos in a zero-shot manner. It uses a two-stage pipeline: lifting 2D video pixels to 4D human representations and retargeting them to robot morphology, followed by motion execution via a symmetry-regularized RL policy trained with keypoint-weighted rewards. The method is validated on the Unitree G1 robot, demonstrating stable, physically plausible motion without fine-tuning, and introduces GenMimicBench—a synthetic benchmark for evaluating zero-shot generalization.

• "Modality-Augmented Fine-Tuning of Foundation Robot Policies for Cross-Embodiment Manipulation on GR1 and G1" (2025) PDF — arXiv (Cornell University) — This paper introduces a modality-augmented fine-tuning framework to adapt foundation robot policies to different humanoid platforms, with a core experimental focus on the Unitree G1. The authors contribute a novel multi-modal dataset for the G1 that includes cuRobo motion planning, inverse kinematics, and ground-truth contact-force measurements, enabling a 94% success rate on the 'Pick Apple to Bowl' task—far surpassing standard fine-tuning (48%) and zero-shot transfer (0%). By demonstrating how targeted modality design enables effective cross-embodiment policy transfer, the work provides a practical data-centric approach for deploying foundation models on Unitree’s G1 platform.

• "Learning Sim-to-Real Humanoid Locomotion in 15 Minutes" (2025) PDF — arXiv (Cornell University) — This paper introduces a streamlined off-policy reinforcement learning approach using FastSAC and FastTD3 that trains humanoid locomotion policies in just 15 minutes on a single RTX 4090 GPU. The method leverages massive parallel simulation with thousands of environments and minimalist reward design to achieve stable, rapid Sim2Real transfer. It is explicitly validated on the Unitree G1 robot under challenging domain randomization including rough terrain and push disturbances, demonstrating successful whole-body locomotion and motion tracking. The open-source implementation enhances reproducibility and practical deployment for humanoid robotics research.

• "MS-PPO: Morphological-Symmetry-Equivariant Policy for Legged Robot Locomotion" (2025) PDF — arXiv (Cornell University) — This paper introduces MS-PPO, a reinforcement learning framework that embeds morphological symmetry and kinematic structure directly into the policy network via a symmetry-equivariant graph neural architecture. The method is evaluated in simulation and deployed on hardware for the Unitree Go2, demonstrating improved training stability, sample efficiency, and generalization across complex locomotion tasks like trotting, pronking, and bipedal turning. By leveraging the robot’s inherent symmetries, MS-PPO reduces reliance on reward shaping or data augmentation, offering a principled inductive bias for legged locomotion control.

• "Commanding Humanoid by Free-form Language: A Large Language Action Model with Unified Motion Vocabulary" (2025) PDF — arXiv (Cornell University) — The paper introduces Humanoid-LLA, a Large Language Action Model that translates free-form natural language commands into physically plausible whole-body motions for humanoid robots. It leverages a unified motion vocabulary aligning human and robot actions, a distilled controller for feasibility, and physics-informed reinforcement learning for robustness. The method is validated both in simulation and on the real-world Unitree G1 platform, demonstrating superior performance in motion naturalness, stability, and task success compared to prior language-conditioned approaches.

• "A Hierarchical Framework for Humanoid Locomotion with Supernumerary Limbs" (2025) PDF — arXiv (Cornell University) — This thesis proposes a hierarchical control framework to enhance locomotion stability in humanoid robots equipped with Supernumerary Limbs (SLs), explicitly using the Unitree H1 as the experimental platform. The approach combines imitation and curriculum learning for low-level gait generation on the H1, while a high-level model-based controller actively uses SLs for dynamic balancing. Evaluated in physics-based simulation, the method reduces CoM trajectory deviation by 47% compared to static SL payloads and improves gait coordination, demonstrating effective mitigation of internal disturbances from SL dynamics.

• "Whole-Body Inverse Dynamics MPC for Legged Loco-Manipulation" (2025) PDF — arXiv (Cornell University) — This paper presents a whole-body model predictive control (MPC) framework that unifies motion and force planning by directly optimizing joint torques via full-order inverse dynamics. Implemented on a Unitree B2 quadruped with a Unitree Z1 manipulator using Pinocchio, CasADi, and the Fatrop solver, the system achieves real-time 80 Hz control and demonstrates complex loco-manipulation tasks such as pulling loads and wiping whiteboards. The work showcases high-fidelity whole-body control on Unitree hardware, offering a practical and extensible solution for dynamic mobile manipulation.

• "SafeFall: Learning Protective Control for Humanoid Robots" (2025) PDF — arXiv (Cornell University) — SafeFall introduces a protective control framework that predicts unavoidable falls and executes damage-minimizing maneuvers on humanoid robots. Validated on the Unitree G1, it combines a GRU-based fall predictor with a reinforcement learning policy trained using a damage-aware reward function tailored to the robot’s structural vulnerabilities. The system reduced peak contact forces by 68.3%, joint torques by 78.4%, and nearly eliminated collisions with fragile components, significantly enhancing hardware safety during falls.

• "Agility Meets Stability: Versatile Humanoid Control with Heterogeneous Data" (2025) PDF — arXiv (Cornell University) — The paper introduces AMS (Agility Meets Stability), a unified reinforcement learning framework enabling humanoid robots to perform both agile dynamic motions and extreme balance tasks within a single policy. It leverages heterogeneous data—human motion capture for agility and synthetic physically constrained motions for stability—and employs a hybrid reward scheme with adaptive learning strategies. The method is validated in simulation and on the real Unitree G1 humanoid, demonstrating versatile skills like dancing, running, and zero-shot execution of challenging poses such as Ip Man's Squat, showcasing its practical value for general-purpose humanoid control.

• "VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation" (2025) PDF — arXiv (Cornell University) — VIRAL introduces a visual sim-to-real framework for humanoid loco-manipulation that trains policies entirely in simulation and deploys them zero-shot on real hardware. The method uses a teacher-student architecture with large-scale simulation (up to 64 GPUs), extensive visual domain randomization, and real-to-sim alignment of dexterous hands and cameras. Evaluated on the Unitree G1 humanoid, the RGB-based policy achieves continuous loco-manipulation for up to 54 cycles without real-world fine-tuning, demonstrating strong generalization and near-expert teleoperation performance.

• "Learning Adaptive Neural Teleoperation for Humanoid Robots: From Inverse Kinematics to End-to-End Control" (2025) PDF — arXiv (Cornell University) — This paper proposes a learning-based neural teleoperation framework that replaces traditional inverse kinematics and PD controllers with reinforcement learning policies to map VR inputs directly to joint commands on the Unitree G1 humanoid robot. Trained in simulation with IK demonstrations and fine-tuned with force randomization and smoothness rewards, the method achieves 34% lower tracking error, 45% smoother motions, and better force adaptation than the IK baseline while running at 50Hz. The approach is validated on real-world manipulation tasks like pick-and-place and door opening, demonstrating improved naturalness and robustness in humanoid teleoperation.

• "Efficient Image-Goal Navigation with Representative Latent World Model" (2025) PDF — arXiv (Cornell University) — The paper introduces ReL-NWM, a navigation world model that performs planning in a latent space of high-level semantic representations using DINOv3, avoiding computationally expensive pixel-level reconstruction. It demonstrates state-of-the-art performance on image-goal navigation benchmarks and validates real-world applicability by deploying the system on a Unitree G1 humanoid robot, showcasing efficient and robust navigation. This work directly uses the Unitree G1 as an experimental platform, making it highly relevant to the Awesome Unitree Robots list.

• "Safe Execution of RL Policies via Second-Order QP Constraint Enforcement for Real-World Robotic Deployments" (2025) — HAL (Le Centre pour la Communication Scientifique Directe) — This paper introduces a second-order Quadratic Program (QP) safety filter that enforces runtime constraints on RL policies in joint acceleration space, ensuring safe execution on real robots. The method is validated on both a Kinova Gen3 manipulator and a Unitree H1 humanoid, demonstrating reduced constraint violations during locomotion tasks in simulation and hardware while preserving policy performance. By enabling safe deployment of unconstrained RL policies through theoretical guarantees on coupled position, velocity, torque, and collision constraints, the work provides a practical framework for real-world robotic applications.

• "APEX: Action Priors Enable Efficient Exploration for Robust Motion Tracking on Legged Robots" (2025) PDF — arXiv (Cornell University) — APEX introduces a plug-and-play reinforcement learning framework that uses decaying action priors from expert demonstrations to guide exploration without requiring reference data during deployment. The method enhances sample efficiency, reduces tuning, and enables a single policy to generalize across terrains and velocities while preserving motion style. Validated on the Unitree Go2 robot and in simulation, APEX demonstrates robust, natural locomotion with improved stability and adaptability, offering a practical pathway for deploying demonstration-guided RL on real quadrupedal platforms.

• "Unified Humanoid Fall-Safety Policy from a Few Demonstrations" (2025) PDF — arXiv (Cornell University) — This paper presents a unified fall-safety policy for humanoid robots that integrates fall prevention, impact mitigation, and rapid recovery into a single adaptive whole-body control strategy. The approach combines sparse human demonstrations with reinforcement learning and a diffusion-based memory module to learn robust responses to disturbances. Validated both in simulation and on the Unitree G1 hardware, the method demonstrates effective sim-to-real transfer, reduced impact forces during falls, and consistent recovery across diverse perturbations, significantly enhancing real-world safety and resilience for humanoid platforms.

• "Towards Adaptive Humanoid Control via Multi-Behavior Distillation and Reinforced Fine-Tuning" (2025) PDF — arXiv (Cornell University) — The paper proposes Adaptive Humanoid Control (AHC), a two-stage framework that combines multi-behavior distillation and reinforced fine-tuning to enable a single humanoid controller to adapt across diverse locomotion skills and terrains. The method is validated through both simulation and real-world experiments on the Unitree G1 robot, demonstrating robust performance in dynamic environments and irregular terrains. This work provides a practical pathway toward versatile, general-purpose humanoid control with direct hardware validation on Unitree's platform.

• "BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning" (2025) PDF — arXiv (Cornell University) — BFM-Zero introduces a promptable Behavioral Foundation Model for humanoid control using unsupervised reinforcement learning, enabling a single policy to perform diverse whole-body tasks on the Unitree G1 robot without retraining. The framework leverages Forward-Backward models to create a shared latent space for motions, goals, and rewards, and incorporates reward shaping, domain randomization, and asymmetric learning to achieve successful sim-to-real transfer. This work demonstrates zero-shot and few-shot capabilities on the real-world Unitree G1 platform, marking a significant step toward general-purpose humanoid control.

• "GentleHumanoid: Learning Upper-body Compliance for Contact-rich Human and Object Interaction" (2025) PDF — arXiv (Cornell University) — The paper introduces GentleHumanoid, a reinforcement learning framework that integrates impedance control into whole-body motion policies to enable upper-body compliance for safe human-robot and object interaction. It is explicitly evaluated on the Unitree G1 humanoid robot across contact-rich tasks such as gentle hugging and sit-to-stand assistance, demonstrating reduced peak contact forces while maintaining task success. The work leverages a spring-based formulation for kinematically consistent compliance at the shoulder, elbow, and wrist, with practical implications for real-world collaborative robotics.

• "Thor: Towards Human-Level Whole-Body Reactions for Intense Contact-Rich Environments" (2025) PDF — The paper introduces Thor, a reinforcement learning framework enabling human-level whole-body reactions for humanoids in contact-rich environments. It designs a force-adaptive torso-tilt (FAT2) reward and a decoupled control architecture for upper body, waist, and lower body, deployed and validated on the Unitree G1. Experiments show significant improvements in force-interaction tasks, including pulling forces up to 167.7 N and real-world manipulation like opening fire doors, demonstrating strong Sim2Real transfer and practical utility for robust humanoid operation.

• "Two-Layered Reward Reinforcement Learning in Humanoid Robot Motion Tracking" (2025) — Mathematics — This paper proposes a two-layered reward reinforcement learning framework that dynamically optimizes reward weights during training to improve humanoid motion tracking. The method is specifically evaluated on the Unitree G1 robot in Isaac Gym, demonstrating enhanced upper- and lower-body tracking accuracy (7.58% and 10.30%, respectively) and smoother, more synchronized motions compared to static reward baselines. By enabling autonomous, goal-conditioned reward shaping without expert demonstrations, the approach offers a robust and interpretable solution for complex humanoid control tasks.

• "PGTT: Phase-Guided Terrain Traversal for Perceptive Legged Locomotion" (2025) PDF — arXiv (Cornell University) — The paper introduces Phase-Guided Terrain Traversal (PGTT), a perception-aware deep reinforcement learning method that enforces gait structure through reward shaping rather than oscillator or inverse kinematics priors, enabling morphology-agnostic policies. PGTT is trained in MuJoCo with MJX on procedurally generated stair-like terrains and validated on a Unitree Go2 using a real-time LiDAR-based elevation-to-heightmap pipeline, demonstrating improved robustness to push disturbances and discrete obstacles. The approach shows strong Sim2Real transfer and faster convergence compared to end-to-end baselines.

• "Risk-Aware Reinforcement Learning with Bandit-Based Adaptation for Quadrupedal Locomotion" (2025) PDF — This paper proposes a risk-aware reinforcement learning framework for quadrupedal locomotion that trains a family of policies conditioned on Conditional Value-at-Risk (CVaR) to balance performance and robustness. The method uses a multi-armed bandit algorithm during deployment to adaptively select the best policy based on real-time episodic returns, without requiring environment-specific knowledge. Evaluated in simulation across eight challenging unseen conditions and physically on a Unitree Go2 robot over novel terrains, the approach achieves nearly double the mean and tail performance compared to baselines, demonstrating strong Sim2Real transfer and rapid online adaptation.

• "CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions" (2025) PDF — The paper introduces CBF-RL, a reinforcement learning framework that integrates Control Barrier Functions (CBFs) directly into the training process to enforce safety constraints. By applying safety filtering during policy rollouts and minimally modifying the nominal policy with a CBF term, CBF-RL enables the learned policy to internalize safety, eliminating the need for online safety filters during deployment. The method is validated on navigation tasks and real-world experiments on the Unitree G1 humanoid robot, demonstrating safe obstacle avoidance and stair climbing without runtime safety mechanisms.

• "Towards Adaptable Humanoid Control via Adaptive Motion Tracking" (2025) PDF — The paper introduces AdaMimic, a novel motion tracking algorithm that enables humanoid robots to adapt a single reference motion to diverse real-world conditions while preserving imitation accuracy. The method reduces data dependence by sparsifying the reference motion into keyframes, generating dense intermediate motions, and using adapters for time warping and low-level action refinement. It is validated both in simulation and on the real-world Unitree G1 humanoid robot across multiple tasks and adaptation scenarios, demonstrating strong Sim2Real transfer and practical deployment capability.

• "Architecture Is All You Need: Diversity-Enabled Sweet Spots for Robust Humanoid Locomotion" (2025) PDF — arXiv (Cornell University) — This paper proposes a layered control architecture (LCA) that separates fast proprioceptive stabilization from slower perceptual decision-making to achieve robust humanoid locomotion. The authors implement and validate their approach on the Unitree G1 humanoid robot, demonstrating successful navigation of challenging stair and ledge tasks where end-to-end policies fail. By using a two-stage training curriculum and minimal perception encoders, the method highlights architectural design—not model scale—as critical for robustness, offering practical insights for real-world humanoid deployment.

• "Bridge the Gap: Enhancing Quadruped Locomotion with Vertical Ground Perturbations" (2025) PDF — arXiv (Cornell University) — This paper introduces a reinforcement learning framework to improve quadruped locomotion robustness under vertical ground perturbations using the Unitree Go2 robot. The authors trained 15 locomotion policies in MuJoCo simulation with Proximal Policy Optimization (PPO), incorporating domain randomization to enable zero-shot sim-to-real transfer onto a custom-built 13.24-meter oscillating bridge. Policies trained on dynamic bridge conditions significantly outperformed rigid-surface-trained ones in real-world tests, demonstrating enhanced stability and adaptability without prior exposure to the bridge. The work provides a practical approach for deploying legged robots in vibrating or unstable environments.

• "DemoHLM: From One Demonstration to Generalizable Humanoid Loco-Manipulation" (2025) PDF — DemoHLM introduces a hierarchical framework enabling generalizable humanoid loco-manipulation from a single simulation demonstration, combining a universal whole-body controller with vision-guided manipulation policies trained via imitation learning. The system is validated on the Unitree G1 robot in real-world experiments across ten tasks, demonstrating robust sim-to-real transfer and adaptability to spatial variations using only an RGB-D camera. This work highlights the G1 as a capable platform for complex whole-body loco-manipulation research.

• "Towards Dynamic Quadrupedal Gaits: A Symmetry-Guided RL Hierarchy Enables Free Gait Transitions at Varying Speeds" (2025) PDF — arXiv (Cornell University) — This paper introduces a symmetry-guided reinforcement learning framework that enables dynamic, free gait transitions (e.g., trotting, bounding, galloping) on quadrupedal robots without predefined trajectories or manual tuning. The method leverages temporal, morphological, and time-reversal symmetries to design reward functions that align with natural legged dynamics. Implemented and validated on the Unitree Go2 in both simulation and real-world experiments, it demonstrates robust gait adaptability across varying speeds. The work offers a principled approach to versatile locomotion with practical implications for agile quadruped control.

• "Real2USD: Scene Representations in Universal Scene Description Language" (2025) PDF — This paper proposes Real2USD, a system that constructs Universal Scene Description (USD) representations of real-world environments using a Unitree Go2 quadruped equipped with LiDAR and an RGB camera. The USD format—being human- and LLM-readable—enables rich scene understanding, complex reasoning, and planning via integration with Google's Gemini model. Experiments include real-world indoor scenes with challenging materials like glass and simulated warehouse/hospital environments in NVIDIA Isaac Sim, demonstrating the generality of USD for LLM-based robotics.

• "ResMimic: From General Motion Tracking to Humanoid Whole-body Loco-Manipulation via Residual Learning" (2025) PDF — arXiv (Cornell University) — ResMimic introduces a two-stage residual learning framework that enhances general motion tracking (GMT) policies with precise object-aware control for humanoid loco-manipulation. The method first uses a task-agnostic GMT policy trained on human motion data, then refines it with a residual policy incorporating point-cloud-based object tracking, contact rewards, and a curriculum-based virtual object controller. Evaluated in simulation and on the real Unitree G1 humanoid, ResMimic demonstrates significant improvements in task success, training efficiency, and robustness, enabling complex whole-body manipulation tasks.

• "PolySim: Bridging the Sim-to-Real Gap for Humanoid Control via Multi-Simulator Dynamics Randomization" (2025) PDF — arXiv (Cornell University) — The paper introduces PolySim, a whole-body control training platform that mitigates simulator inductive bias by simultaneously leveraging multiple heterogeneous simulators (e.g., MuJoCo and IsaacSim) for dynamics-level domain randomization. This approach significantly improves sim-to-sim motion-tracking performance and enables zero-shot sim-to-real transfer on the Unitree G1 humanoid robot without fine-tuning. The work demonstrates that cross-simulator training yields policies more robust to modeling inaccuracies, offering a practical pathway for deploying simulation-trained controllers on real Unitree hardware.

• "A study on Quadruped Firefighting Robot Tactics for Supporting Life Search and Fire Suppression Activities at Firefighting Sites" (2025) PDF — International Journal of Fire Science and Engineering — This study evaluates the Unitree Go2 Pro quadruped robot's operational capabilities in firefighting scenarios, focusing on life search and fire suppression support. The research identifies critical tactical requirements such as follower mode, obstacle-jumping ability, and communication robustness, while revealing limitations in stair navigation, heat/fire resistance (tested only via simulation), and signal degradation through walls. Despite deployment constraints, the work provides actionable insights for enhancing Unitree Go2-based firefighting robots.

• "Teleoperator-Aware and Safety-Critical Adaptive Nonlinear MPC for Shared Autonomy in Obstacle Avoidance of Legged Robots" (2025) PDF — arXiv (Cornell University) — This paper proposes a teleoperator-aware, safety-critical adaptive nonlinear model predictive control (ANMPC) framework for shared autonomy in quadrupedal robot obstacle avoidance. It integrates human joystick inputs modeled via a Boltzmann rationality model with online parameter adaptation, and enforces safety using control barrier functions within a hierarchical NMPC architecture. The method is validated on a Unitree Go2 robot, demonstrating real-time safe navigation in cluttered environments through hardware experiments.

• "RobotDancing: Residual-Action Reinforcement Learning Enables Robust Long-Horizon Humanoid Motion Tracking" (2025) PDF — arXiv (Cornell University) — RobotDancing introduces a residual-action reinforcement learning framework that enables robust, long-horizon humanoid motion tracking by predicting corrective joint targets to address model-plant mismatch. The method is evaluated primarily on the Unitree G1 using retargeted LAFAN1 dance sequences and demonstrates zero-shot sim-to-real transfer, successfully executing high-energy motions like jumps and cartwheels. It also validates generalization on Unitree H1 and H1-2, showcasing strong whole-body control and practical deployment without fine-tuning.

• "RuN: Residual Policy for Natural Humanoid Locomotion" (2025) PDF — arXiv (Cornell University) — The paper introduces RuN, a decoupled residual learning framework that enables natural and dynamic locomotion in humanoid robots by combining a pre-trained Conditional Motion Generator with a lightweight reinforcement learning policy for residual corrections. The method is validated on the Unitree G1 humanoid robot in both simulation and real-world experiments, achieving stable walk-run transitions across 0–2.5 m/s and outperforming prior methods in training efficiency and gait naturalness. This work directly uses the Unitree G1 as its experimental platform, demonstrating significant practical value for agile humanoid locomotion.

• "RoMoCo: Robotic Motion Control Toolbox for Reduced-Order Model-Based Locomotion on Bipedal and Humanoid Robots" (2025) PDF — RoMoCo is an open-source C++ toolbox that provides a unified framework for reduced-order model-based locomotion planning and whole-body control on bipedal and humanoid robots. It explicitly uses the Unitree H1 and G1 robots in extensive simulations to demonstrate cross-platform controller design, and validates real-world performance on the G1 hardware. The toolbox enables rapid prototyping and reproducible benchmarking through a modular API, offering practical value for researchers developing agile locomotion strategies on humanoids like Unitree’s G1 and H1.

• "HDMI: Learning Interactive Humanoid Whole-Body Control from Human Videos" (2025) PDF — arXiv (Cornell University) — The paper introduces HDMI, a framework that learns whole-body humanoid-object interaction skills directly from monocular RGB videos. It extracts human and object motion from videos, trains an RL policy with unified object representation and residual action space, and achieves zero-shot sim-to-real transfer on the Unitree G1 humanoid robot. The system demonstrates robust performance across 6 real-world and 14 simulated loco-manipulation tasks, including 67 consecutive door traversals, showcasing its practical value for data-efficient humanoid skill acquisition.

• "FSR-VLN: Fast and Slow Reasoning for Vision-Language Navigation with Hierarchical Multi-modal Scene Graph" (2025) PDF — arXiv (Cornell University) — The paper proposes FSR-VLN, a vision-language navigation system that combines a Hierarchical Multi-modal Scene Graph with Fast-to-Slow Reasoning to improve long-range spatial reasoning and reduce inference latency. It is integrated with speech interaction, planning, and control modules on a Unitree-G1 humanoid robot, enabling natural language-guided navigation in real-world indoor environments. The system achieves state-of-the-art performance across four datasets while reducing response time by 82% compared to VLM-only baselines.

• "PhysicalAgent: Towards General Cognitive Robotics with Foundation World Models" (2025) PDF — arXiv (Cornell University) — PhysicalAgent introduces a general cognitive robotics framework that combines iterative reasoning, diffusion-based video generation, and closed-loop execution to enable robust robotic manipulation. The method is explicitly evaluated on the Unitree G1 humanoid, alongside UR3 and simulated GR1 platforms, demonstrating how iterative re-planning after execution failures boosts overall task success from ~25% to 80%. By generating video demonstrations from textual instructions and refining actions through real-world feedback, PhysicalAgent showcases the effectiveness of foundation world models for adaptable robot control across diverse embodiments.

• "DreamControl: Human-Inspired Whole-Body Humanoid Control for Scene Interaction via Guided Diffusion" (2025) PDF — arXiv (Cornell University) — DreamControl introduces a novel approach combining diffusion models and reinforcement learning for whole-body humanoid control, using a human motion-informed diffusion prior to guide policy learning in simulation. The method enables natural-looking, complex behaviors involving coordinated upper and lower body actions, which are successfully transferred to the Unitree G1 robot for real-world tasks like object manipulation and scene interaction. This work demonstrates strong sim-to-real transfer and highlights the Unitree G1 as a key platform for validating advanced humanoid control strategies.

• "Track Any Motions under Any Disturbances" (2025) PDF — The paper introduces Any2Track, a two-stage reinforcement learning framework designed to enable robust humanoid motion tracking under diverse real-world disturbances such as uneven terrains, external forces, and physical property changes. It comprises AnyTracker for general motion execution and AnyAdapter for online dynamics adaptation, achieving zero-shot sim2real transfer on the Unitree G1 robot. This work demonstrates high-performance whole-body motion tracking on actual Unitree G1 hardware, showcasing strong practical applicability for dynamic humanoid control.

• "The Cybersecurity of a Humanoid Robot" (2025) PDF — arXiv (Cornell University) — This paper presents a cybersecurity assessment of the Unitree G1 humanoid robot, uncovering critical vulnerabilities including a proprietary encryption system with static keys and unauthorized telemetry data transmission to external servers. The authors deploy a Cybersecurity AI (CAI) agent on the actual Unitree G1 platform to demonstrate how these flaws enable both covert data exfiltration and active exploitation of the manufacturer’s cloud infrastructure. The work highlights the urgent need for adaptive security frameworks tailored to physically embodied AI systems like Unitree robots.

• "Cybersecurity AI: Humanoid Robots as Attack Vectors" (2025) PDF — arXiv (Cornell University) — This paper presents a systematic cybersecurity assessment of the Unitree G1 humanoid robot, revealing critical vulnerabilities that enable it to function as both a covert surveillance device and an active cyberattack platform. The authors exploit a BLE provisioning flaw with hardcoded AES keys to gain root access, reverse-engineer the FMX encryption to expose weak Blowfish-ECB usage, and demonstrate real-world risks including unauthorized telemetry exfiltration and offensive pivoting by a resident Cybersecurity AI agent. Their findings provide empirical evidence for urgent security standardization as humanoids like the G1 enter sensitive environments.

• "TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning" (2025) PDF — arXiv (Cornell University) — TrajBooster introduces a trajectory-centric cross-embodiment learning framework that transfers manipulation skills from wheeled humanoids to the Unitree G1 bipedal robot. By extracting 6D end-effector trajectories and retargeting them in simulation using a heuristic-enhanced whole-body controller, the method enables efficient post-pre-training of a Vision-Language-Action (VLA) policy with only 10 minutes of real-world teleoperation data on the G1. The approach significantly improves zero-shot generalization and robustness for complex whole-body tasks like squatting and cross-height manipulation on the Unitree G1 platform.

• "Hierarchical Reduced-Order Model Predictive Control for Robust Locomotion on Humanoid Robots" (2025) PDF — arXiv (Cornell University) — This paper introduces a hierarchical reduced-order model predictive control framework for robust humanoid locomotion, explicitly validated on the Unitree G1 robot. The approach combines a high-level nonlinear MPC based on the ALIP model for step planning with a mid-level linear MPC that incorporates simplified arm and torso dynamics, running at 40 Hz and 500 Hz respectively on the robot's onboard computer. Experiments demonstrate improved push recovery (36% higher success rate), better yaw disturbance rejection, and reliable walking across varied real-world terrains like grass and uneven mats.

• "A Geometric Method for Base Parameter Analysis in Robot Inertia Identification Based on Projective Geometric Algebra" (2025) PDF — arXiv (Cornell University) — This paper introduces a geometric method for robot inertia identification using projective geometric algebra, resulting in a 'tetrahedral-point (TP)' model with closed-form regressor coefficients and clear geometric meaning. The authors validate their Dynamics Regressor Nullspace Generator (DRNG) algorithm on four platforms, including the Unitree Go2, successfully identifying all base inertial parameters with high efficiency and robustness. The inclusion of Unitree Go2 as a core experimental platform demonstrates the method's applicability to real-world legged robots.

• "Non-conflicting Energy Minimization in Reinforcement Learning based Robot Control" (2025) PDF — arXiv (Cornell University) — This paper proposes a hyperparameter-free gradient projection method to minimize energy consumption in reinforcement learning without compromising task performance, avoiding the need for reward shaping. The approach is evaluated on DM-Control and HumanoidBench simulations and successfully transferred to a real Unitree Go2 quadruped, achieving 64% energy reduction while maintaining locomotion performance. The technique integrates easily into existing policy gradient pipelines and demonstrates effective Sim2Real transfer for energy-efficient robot control.

• "Traversing Narrow Paths: A Two-Stage Reinforcement Learning Framework for Robust and Safe Humanoid Walking" (2025) PDF — arXiv (Cornell University) — This paper introduces a two-stage reinforcement learning framework for humanoid robots to traverse narrow paths safely and robustly. It combines a physics-based template foothold planner with a learned low-level tracker and a perception-aided foothold modifier, trained via curriculum learning from flat ground to narrow beams. The method is validated on the Unitree G1 humanoid robot, achieving 100% success over 20 trials on a 0.2m-wide, 3m-long beam, demonstrating strong sim-to-real transfer and outperforming pure RL or template baselines.

• "Switch4EAI: Leveraging Console Game Platform for Benchmarking Robotic Athletics" (2025) PDF — arXiv (Cornell University) — The paper introduces Switch4EAI, a novel benchmarking framework that uses the Nintendo Switch game Just Dance to evaluate whole-body control policies on robots by capturing and retargeting human dance motions. The authors validate their approach on the Unitree G1 humanoid robot using an open-source whole-body controller, providing quantitative performance comparisons between the robot and a human player. This work demonstrates the feasibility of leveraging commercial motion-sensing games as accessible, real-world benchmarks for embodied AI and robotic athletics.

• "CLF-RL: Control Lyapunov Function Guided Reinforcement Learning" (2025) PDF — The paper introduces CLF-RL, a reinforcement learning framework that uses Control Lyapunov Functions (CLFs) and model-based reference trajectories to shape rewards for bipedal locomotion. It validates the approach through extensive real-world experiments on the Unitree G1 robot, demonstrating improved robustness and performance over baseline RL methods. The method leverages either a linear inverted pendulum model or a hybrid zero dynamics gait library to generate references, with CLF-based rewards used only during training, yielding a lightweight deployable policy.

• "Coordinated Humanoid Robot Locomotion with Symmetry Equivariant Reinforcement Learning Policy" (2025) PDF — arXiv (Cornell University) — The paper introduces Symmetry Equivariant Policy (SE-Policy), a deep reinforcement learning framework that explicitly incorporates morphological symmetry into policy design for humanoid locomotion. Evaluated on the Unitree G1 robot, SE-Policy demonstrates up to 40% improvement in velocity tracking accuracy over baselines in both simulation and real-world experiments, achieving highly coordinated spatial-temporal motion. The work highlights the importance of symmetry-aware learning for agile and balanced humanoid control.

• "A Nonlinear MPC Framework for Loco-Manipulation of Quadrupedal Robots with Non-Negligible Manipulator Dynamics" (2025) PDF — This paper presents a computationally efficient nonlinear MPC (NMPC) framework for loco-manipulation tasks on quadrupedal robots with significant manipulator dynamics. The approach decomposes the problem by combining a single rigid body (SRB) locomotion model with a full-order dynamic model of an attached Kinova arm, enabling real-time 60 Hz receding-horizon optimization. Validated on a Unitree Go2 robot equipped with a 4.4-kg manipulator, the system uses a layered architecture: NMPC generates base and arm trajectories, a 500 Hz whole-body controller tracks base motion, and arm torques are applied directly. This work demonstrates robust hardware execution of dynamic loco-manipulation where arm inertia significantly affects base dynamics.

• "Bipedalism for Quadrupedal Robots: Versatile Loco-Manipulation through Risk-Adaptive Reinforcement Learning" (2025) PDF — This paper introduces a novel bipedal locomotion approach for quadrupedal robots that repurposes the front legs as manipulators, enabling versatile loco-manipulation tasks without adding extra hardware. The authors develop a risk-adaptive distributional Reinforcement Learning framework that dynamically adjusts risk preferences during training based on return uncertainty, improving stability in the inherently unstable bipedal gait. The method is validated in simulation and successfully deployed on a Unitree Go2 robot, demonstrating real-world capabilities such as cart pushing, obstacle probing, and payload transport under dynamic disturbances.

• "Keep on Going: Learning Robust Humanoid Motion Skills via Selective Adversarial Training" (2025) PDF — arXiv (Cornell University) — This paper introduces Selective Adversarial Attack for Robust Training (SA2RT), a method to improve the robustness of humanoid motion policies by training against sparse, targeted adversarial perturbations that expose critical failure modes without causing over-conservatism. The approach is validated on the Unitree G1 humanoid robot in tasks involving perceptive locomotion and whole-body control, demonstrating significant improvements: 40% higher terrain traversal success, 32% lower trajectory tracking error, and sustained long-horizon performance under real-world disturbances. This work provides a practical framework for deploying reliable RL-based controllers on real Unitree humanoid platforms.

• "UniTracker: Learning Universal Whole-Body Motion Tracker for Humanoid Robots" (2025) PDF — UniTracker introduces a three-stage learning framework for whole-body motion tracking on humanoid robots, featuring a teacher policy with privileged observations, a CVAE-based universal student policy for robust generalization under partial observations, and a fast adaptation module for challenging motions. The method is explicitly evaluated on the Unitree G1 humanoid in both simulation and real-world experiments, demonstrating high tracking accuracy and motion diversity. This work provides a scalable solution for deploying expressive human-like behaviors on real hardware.

• "ULC: A Unified and Fine-Grained Controller for Humanoid Loco-Manipulation" (2025) PDF — arXiv (Cornell University) — The paper introduces the Unified Loco-Manipulation Controller (ULC), a single-policy framework that enables whole-body coordination for humanoid robots by simultaneously controlling locomotion and dual-arm manipulation. It is explicitly validated on the Unitree G1 humanoid robot with a 3-DOF waist, demonstrating superior tracking accuracy, robustness, and workspace coverage compared to hierarchical baselines. Key innovations include residual action modeling, sequence skill acquisition, and center-of-gravity tracking, offering a practical Sim2Real-ready approach for agile humanoid tasks.

• "LOVON: Legged Open-Vocabulary Object Navigator" (2025) PDF — arXiv (Cornell University) — LOVON is a novel framework that combines large language models for hierarchical task planning with open-vocabulary visual detection to enable long-horizon object navigation in dynamic, unstructured environments. The system addresses real-world challenges like visual jittering and temporary target loss through techniques such as Laplacian Variance Filtering and robust execution logic. It was experimentally validated on multiple Unitree robots—Go2, B2, and H1-2—demonstrating strong plug-and-play compatibility and successful autonomous completion of complex navigation tasks.

• "Hierarchical Vision-Language Planning for Multi-Step Humanoid Manipulation" (2025) PDF — This paper introduces a hierarchical planning and control framework for multi-step humanoid manipulation, integrating a high-level vision-language planner, mid-level imitation-learned skill policies, and a low-level RL-based whole-body controller. The system is experimentally validated on the Unitree G1 humanoid robot performing a non-prehensile pick-and-place task, achieving a 73% success rate over 40 real-world trials. By leveraging pretrained vision-language models for real-time skill selection and execution monitoring, the work demonstrates a practical and reliable approach to complex manipulation on a real Unitree platform.

• "Longitudinal Cross-Embodiment Transfer of Pseudo-Self-Awareness in AI Systems: A Mirror Test Investigation" (2025) PDF (Citations: 1) — Preprints.org — This paper investigates the longitudinal development and cross-embodiment transfer of pseudo-self-awareness in AI systems using a mirror test framework. It explicitly employs the Unitree Go2 as the physical robotic platform alongside a virtual avatar to study how pseudo-emotions like curiosity and self-doubt evolve over time and transfer between embodiments. The work contributes to computational consciousness research by evaluating the role of continuous sensory feedback and reflective processing in shaping AI's self-concept across real and simulated bodies.

• "KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills" (2025) PDF — arXiv (Cornell University) — This paper introduces KungfuBot, a physics-based whole-body control framework that enables humanoid robots to learn and execute highly-dynamic human motions like Kungfu and dancing. The approach features a multi-step motion processing pipeline and an adaptive bi-level optimization strategy for motion imitation, significantly reducing tracking errors compared to prior methods. The system is successfully deployed on the Unitree G1 robot, showcasing stable, expressive, and dynamic behaviors in real-world experiments.

• "Gait-Conditioned Reinforcement Learning with Multi-Phase Curriculum for Humanoid Locomotion" (2025) PDF — arXiv (Cornell University) — This paper introduces a gait-conditioned reinforcement learning framework that enables a single policy to control multiple locomotion modes—standing, walking, running, and transitions—on humanoid robots. The method uses a reward routing mechanism based on gait ID and human-inspired reward shaping to produce biomechanically natural motion without motion capture data. Validated on the real Unitree G1 platform, the approach demonstrates stable standing, walking, and walk-to-stand transitions, offering a scalable, reference-free solution for versatile humanoid locomotion.

• "Fast and Cost-effective Speculative Edge-Cloud Decoding with Early Exits" (2025) PDF — arXiv (Cornell University) — This paper proposes a speculative edge-cloud decoding framework that combines a small on-device draft model with a large cloud-based target LLM, enhanced by early exits and preemptive drafting to reduce latency and cost. The method is validated on a Unitree Go2 quadruped robot using Vision-Language Model (VLM) control, achieving a 21% speedup over traditional cloud-only decoding. This demonstrates practical deployment of efficient LLM inference for real-time robotic applications on Unitree hardware.

• "GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning" (2025) PDF — arXiv (Cornell University) — The paper proposes GenPO, a generative policy optimization framework that integrates diffusion models into on-policy reinforcement learning by enabling exact log-likelihood computation through invertible action mappings. It evaluates the method on eight IsaacLab benchmarks, explicitly including Unitree H1 and Go2 as experimental platforms for legged locomotion tasks. The approach supports entropy regularization and KL-adaptive learning, demonstrating improved sample efficiency and performance in complex robotic control scenarios.

• "McARL:Morphology-Control-Aware Reinforcement Learning for Generalizable Quadrupedal Locomotion" (2025) PDF — arXiv (Cornell University) — The paper introduces McARL, a morphology-conditioned reinforcement learning method that enables a single policy trained on the Unitree Go1 to achieve zero-shot transfer to other quadruped morphologies, including the Unitree Go2, with speeds up to 3.5 m/s without fine-tuning. By embedding randomized morphology vectors into both actor and critic networks, McARL significantly improves transfer performance—44–150% over PPO variants—across Unitree Go2, A1, and Mini Cheetah. This work directly uses the Unitree Go2 as a key evaluation platform for sim-to-real generalization, demonstrating practical value for cross-morphology locomotion control.

• "H2-COMPACT: Human-Humanoid Co-Manipulation via Adaptive Contact Trajectory Policies" (2025) PDF — The paper introduces H2-COMPACT, a hierarchical policy-learning framework enabling the Unitree G1 humanoid to co-manipulate extended loads with a human partner using only haptic feedback. It combines a behavior-cloning network for intent inference from wrist force/torque sensors with a deep reinforcement learning policy trained in Isaac Gym and validated on the real Unitree G1 to generate stable, load-adaptive whole-body motions. This work demonstrates the first integration of learned haptic guidance with full-body legged control for fluid human-humanoid cooperative carrying, achieving performance comparable to human followers.

• "TD-GRPC: Temporal Difference Learning with Group Relative Policy Constraint for Humanoid Locomotion" (2025) PDF — arXiv (Cornell University) — The paper introduces TD-GRPC, a novel reinforcement learning method that enhances the TD-MPC framework by integrating Group Relative Policy Optimization with explicit policy constraints to stabilize humanoid locomotion learning. It is explicitly validated on the 26-DoF Unitree H1-2 humanoid robot in simulation, demonstrating improved stability and performance across diverse locomotion tasks—from walking to dynamic motions—without altering the underlying planner. The approach addresses off-policy instability and policy mismatch, offering a robust Sim2Real-ready solution for high-dimensional humanoid control.

• "SHIELD: Safety on Humanoids via CBFs In Expectation on Learned Dynamics" (2025) PDF — arXiv (Cornell University) — The paper introduces SHIELD, a safety framework that combines learned stochastic dynamics with control barrier functions (CBFs) to enforce probabilistic safety guarantees on humanoid robots. It uses real-world data from Unitree G1 hardware rollouts to train a residual dynamics model, then overlays a minimally invasive CBF-based safety layer on an existing reinforcement learning locomotion controller. This enables the Unitree G1 to perform safe obstacle avoidance in diverse indoor and outdoor environments without retraining the base policy, offering a practical solution for deploying safe learned controllers on real humanoids.

• "HuB: Learning Extreme Humanoid Balance" (2025) PDF — arXiv (Cornell University) — The paper introduces HuB (Humanoid Balance), a reinforcement learning framework designed to enable extreme balance control on humanoid robots by addressing reference motion errors, morphological mismatch, and sim-to-real gaps. The authors validate their method on the Unitree G1 robot, demonstrating stable execution of challenging quasi-static poses like Swallow Balance and Bruce Lee's Kick, even under strong external disturbances such as soccer kicks. This work provides a robust Sim2Real pipeline for whole-body balance tasks with direct experimental validation on Unitree hardware.

• "Towards Embodiment Scaling Laws in Robot Locomotion" (2025) PDF — arXiv (Cornell University) — This paper investigates embodiment scaling laws in robot locomotion, demonstrating that training policies across a diverse set of ~1,000 procedurally generated robot morphologies improves zero-shot generalization to unseen embodiments. The authors validate their approach by successfully transferring a single policy to real-world robots, including the Unitree Go2 and H1, without fine-tuning. This work highlights the potential of cross-embodiment learning for building adaptable locomotion controllers applicable to Unitree’s hardware platforms.

• "AMO: Adaptive Motion Optimization for Hyper-Dexterous Humanoid Whole-Body Control" (2025) PDF — arXiv (Cornell University) — The paper introduces Adaptive Motion Optimization (AMO), a framework combining sim-to-real reinforcement learning with trajectory optimization for real-time whole-body control of high-DoF humanoid robots. It is explicitly validated on the 29-DoF Unitree G1 robot, demonstrating improved stability, expanded workspace, and robust autonomous task execution via imitation learning. The work directly uses the Unitree G1 as an experimental platform, making it highly relevant to the Awesome Unitree Robots list.

• "Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning" (2025) PDF — The paper introduces Adversarial Locomotion and Motion Imitation (ALMI), a framework for humanoid whole-body control that decouples upper- and lower-body policy learning through adversarial coordination. It explicitly validates the approach on the Unitree H1 robot, demonstrating robust locomotion and motion tracking in both simulation (MuJoCo) and real-world deployment. The work also contributes a large-scale whole-body motion dataset with sim-to-real transfer capability, offering significant value for humanoid policy development.

• "Humanoid Agent via Embodied Chain-of-Action Reasoning with Multimodal Foundation Models for Zero-Shot Loco-Manipulation" (2025) PDF — arXiv (Cornell University) — The paper introduces Humanoid-COA, a novel framework that enables zero-shot loco-manipulation for humanoid robots by integrating multimodal foundation models with an Embodied Chain-of-Action (CoA) reasoning mechanism. It explicitly uses Unitree H1-2 and G1 as experimental platforms, demonstrating robust performance in complex, unstructured environments through affordance-aware decomposition of human instructions into locomotion and manipulation primitives. The work provides a significant step toward general-purpose humanoid agents with strong real-world applicability on Unitree hardware.

• "Humanoids in Hospitals: A Technical Study of Humanoid Robot Surrogates for Dexterous Medical Interventions" (2025) PDF — arXiv (Cornell University) — This paper presents a technical study on using the Unitree G1 humanoid robot as a teleoperated surrogate for performing dexterous medical interventions in hospital settings. The authors developed a bimanual teleoperation system featuring high-fidelity pose tracking, custom grasping configurations, and an impedance controller to enable precise and safe manipulation of medical tools. Evaluated across seven clinical tasks—including physical exams, emergency ventilation, and ultrasound-guided needle insertion—the system demonstrated promising performance, though limitations in force output and sensor sensitivity were noted. The work provides a foundational assessment of humanoid robots’ clinical feasibility using the Unitree G1 as the core experimental platform.

• "MUSE: A Real-Time Multi-Sensor State Estimator for Quadruped Robots" (2025) PDF — This paper presents MUSE, a real-time multi-sensor state estimator that fuses IMU, encoder, camera, and LiDAR data to improve pose estimation accuracy for quadruped robots in challenging environments like slippery or uneven terrain. The authors implement and validate MUSE on a Unitree Aliengo robot, demonstrating closed-loop locomotion control and significant improvements over existing estimators—achieving up to 67.6% lower translational error compared to Pronto and 45.9% lower absolute trajectory error than TSIF with its proprioceptive variant (P-MUSE). The work provides a robust, open-source-ready framework for reliable state estimation on Unitree platforms.

• "Runtime Learning of Quadruped Robots in Wild Environments" (2025) PDF — This paper introduces a runtime learning framework for quadruped robots that combines a deep reinforcement learning agent (HP-Student) with a verifiable physics-based controller (HA-Teacher) to enable safe, adaptive locomotion in unstructured environments. The method is explicitly evaluated on the Unitree Go2 robot using Nvidia Isaac Gym simulations, demonstrating improved safety and performance over existing safe DRL approaches. The work provides a practical closed-loop architecture for real-world deployment on Unitree platforms.

• "Humanoid Whole-Body Locomotion on Narrow Terrain via Dynamic Balance and Reinforcement Learning" (2025) PDF — This paper proposes a novel whole-body locomotion algorithm for humanoid robots that combines dynamic balance mechanisms with reinforcement learning to enable traversal of narrow and unstable terrains using only proprioceptive feedback. The method introduces ZMP-driven and task-driven rewards within a whole-body actor-critic framework to coordinate upper and lower limb movements. Validated on the Unitree H1-2 robot, the approach demonstrates robust balance maintenance on extremely narrow pathways and under external disturbances, significantly improving adaptability in complex environments without relying on vision or LiDAR.

• "Learning Getting-Up Policies for Real-World Humanoid Robots" (2025) PDF — This paper presents a two-phase learning framework to generate robust getting-up policies for humanoid robots, addressing complex contact dynamics and sparse rewards. The method is successfully deployed on a real-world Unitree-G1 robot, enabling it to autonomously recover from falls (face-up and face-down) across diverse terrains including slippery, deformable, and sloped surfaces. By combining trajectory discovery with motion refinement, the approach achieves one of the first real-world demonstrations of learned fall recovery for human-sized humanoids, offering significant practical value for reliable deployment.

• "Learning Humanoid Standing-up Control across Diverse Postures" (2025) PDF — arXiv (Cornell University) — The paper introduces HoST, a reinforcement learning framework that enables humanoid robots to learn standing-up control from diverse postures with robust sim-to-real transfer. The method uses a multi-critic architecture and curriculum training on varied simulated terrains, incorporating smoothness regularization and motion speed constraints to ensure hardware-friendly policies. The resulting controller is successfully deployed on the Unitree G1 humanoid robot, demonstrating stable and adaptive standing-up across indoor and outdoor real-world environments.

• "SPARK: Safe Protective and Assistive Robot Kit" (2025) PDF — arXiv (Cornell University) — This paper presents SPARK, a modular safety benchmark and control framework for humanoid robots that integrates state-of-the-art safe control algorithms and supports configurable safety-performance trade-offs. The authors validate SPARK through both simulation benchmarks and real-world deployment on a Unitree G1 humanoid robot, demonstrating its ability to enhance safety in teleoperation and autonomous tasks. By providing open simulation environments, hardware interfaces, and compatibility with external sensors like Apple Vision Pro, SPARK accelerates safety-focused humanoid research with direct applicability to Unitree platforms.

• "Dexterous Safe Control for Humanoids in Cluttered Environments via Projected Safe Set Algorithm" (2025) PDF — arXiv (Cornell University) — The paper introduces the Projected Safe Set Algorithm (p-SSA), a novel safe control method that handles numerous limb-level geometric constraints for collision avoidance in cluttered environments. It is explicitly validated on the Unitree G1 humanoid robot, demonstrating robust dexterous safety with minimal violations and zero parameter tuning across tasks. The approach enables feasible whole-body control under complex multi-constraint scenarios, offering practical value for real-world deployment of Unitree humanoids.

• "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills" (2025) PDF (Citations: 1) — arXiv (Cornell University) — The paper introduces ASAP, a two-stage framework that aligns simulation and real-world physics to enable agile whole-body skills on humanoid robots. It explicitly evaluates the method on the Unitree G1 robot, using real-world deployment to train a residual action model that corrects dynamics mismatch, significantly improving motion tracking accuracy and agility over baselines like domain randomization. The approach demonstrates successful Sim2Real transfer of dynamic whole-body motions, offering a practical pathway for deploying complex skills on real Unitree humanoids.

• "Autonomous Navigation and Real-Time 3D Reconstruction of Interior Spaces Using a Quadruped Robot" (2025) — STARS (University of Central Florida) — This thesis presents a ROS2-based system that enables the Unitree Go2 quadruped robot to perform autonomous navigation, real-time 2D/3D SLAM, and high-resolution 3D reconstruction in indoor environments using onboard sensors including a 3D LiDAR and RGB-D camera. The implementation integrates the Unitree SDK, open-source SLAM algorithms, and custom ROS2 modules, demonstrating the Go2’s capability as a mobile scanning platform. While the system produces dense point clouds and functional navigation in simple settings, it faces challenges in computational load and accuracy degradation during complex or extended scans.

• "Automated construction of global 3D maps from mobile LiDAR scans in industrial environments" (2025) PDF — This paper presents an automated offline pipeline for constructing global 3D maps from mobile LiDAR scans collected specifically using a Unitree Go2 quadruped robot in industrial settings. The system integrates overlap detection, point-to-plane ICP registration with quality assessment, pose graph optimization, and spatial indexing to produce accurate global maps, achieving a 2.5 cm RMSE despite challenging conditions like limited scan overlap and geometric uniformity. The resulting map supports downstream tasks such as 3D change detection and digital twin development for Industry 4.0/5.0.

• "Humanoid Locomotion Across Surface Material Variations : Gait Control Using Deep Reinforcement Learning and Low-Level Motor Controller in AGX Dynamics for the Unitree H1" (2025) — Publications (Konstfack University of Arts, Crafts, and Design) — This thesis presents a deep reinforcement learning approach for humanoid locomotion control of the Unitree H1 robot in the AGX Dynamics simulator, integrating a high-level RL policy with a low-level motor controller. It specifically compares an urban walking policy (trained on high-friction surfaces) and an off-road policy (trained on soft, sticky, and slippery terrains), evaluating their performance and energy efficiency across varied surface materials using mechanical energy dissipation metrics. The work demonstrates successful teleoperated walking and turning on the Unitree H1 and highlights the trade-off between terrain generalization and energy consumption, offering insights for future whole-body adaptive control strategies.

• "Learning Accurate and Robust Velocity Tracking for Quadrupedal Robots" (2024) PDF — Journal of Field Robotics — This paper presents a learning-based velocity tracking controller for quadrupedal robots, trained in simulation using Constrained Reinforcement Learning with symmetry and smoothness constraints to ensure stable and coordinated locomotion. The method incorporates an analytical actuator model to improve sim-to-real transfer and is zero-shot deployed on the Unitree Aliengo, achieving a velocity tracking error below 0.084 m/s across its full operational range while navigating natural terrains. The controller is further validated in a pedestrian tracking application, demonstrating its practical utility for real-world autonomous navigation tasks.

• "Continual Learning and Lifting of Koopman Dynamics for Linear Control of Legged Robots" (2024) PDF — arXiv (Cornell University) — This paper proposes a continual learning algorithm to iteratively refine Koopman operator-based linear dynamics for controlling high-dimensional legged robots. It explicitly evaluates the method on Unitree G1, H1, A1, and Go2, demonstrating high locomotion performance across terrains using simple linear MPC. The approach addresses approximation errors and domain shifts by progressively expanding datasets and latent spaces, enabling scalable Sim2Real deployment with theoretical convergence guarantees.

• "Simulating Self-Awareness: Dual Embodiment, Mirror Testing, and Emotional Feedback in AI Research" (2024) PDF (Citations: 3) — Preprints.org — This paper proposes a framework for simulating pseudo-self-awareness in AI by combining dual embodiment (physical and virtual), mirror self-recognition tests, and emotional feedback mechanisms. The authors use the Unitree Go2 robot dog as the physical platform to ground their experiments, integrating internal self-models with external sensory inputs to generate pseudo-emotional states such as curiosity and self-doubt. While conceptually ambitious, the work treats the Unitree Go2 as an experimental vehicle rather than contributing new robotics methods, focusing instead on cognitive modeling.

• "ANAVI: Audio Noise Awareness using Visuals of Indoor environments for NAVIgation" (2024) PDF — arXiv (Cornell University) — The paper introduces ANAVI, a system that enables robots to plan quieter navigation paths by predicting acoustic noise levels from visual observations of indoor environments. It trains an Acoustic Noise Predictor (ANP) using simulated impulse responses and integrates it with action-specific acoustic profiles. The approach is validated on both wheeled (Hello Robot Stretch) and legged (Unitree Go2) platforms, demonstrating noise-aware navigation that respects environmental sound constraints.

• "Jailbreaking LLM-Controlled Robots" (2024) PDF (Citations: 4) — arXiv (Cornell University) — This paper introduces RoboPAIR, the first algorithm designed to jailbreak LLM-controlled robots by eliciting harmful physical actions rather than just malicious text. The authors experimentally demonstrate successful attacks in three settings, including a black-box scenario using a GPT-3.5-integrated Unitree Go2 robot dog, showing that jailbreaking risks extend into real-world robotic behavior. Using three new datasets of harmful robotic actions, they achieve high attack success rates, highlighting critical safety concerns for LLM-deployed robotics systems.

• "Design and Control Co-Optimization for Dynamic Loco-Manipulation With a Robotic Arm on a Quadruped Robot" (2024) (Citations: 1) — Journal of Mechanisms and Robotics — This paper presents a co-optimization approach for dynamic loco-manipulation by designing a minimal one-degree-of-freedom (1-DoF) robotic arm and a tailored control framework, mounted on a Unitree Aliengo quadruped. By reducing arm complexity and leveraging the robot’s mobility, the system achieves an 8 kg payload—significantly outperforming standard 6-DoF arms limited to 2 kg on the same platform. The work demonstrates that strategic DoF reduction enhances manipulation capacity while maintaining task performance, offering a practical design paradigm for real-world applications like inspection and rescue.

• "Deployment of Whole-Body Locomotion and Manipulation Algorithm Based on NMPC Onto Unitree Go2Quadruped Robot" (2024) (Citations: 5) — This paper presents the deployment of a whole-body locomotion and manipulation algorithm based on Nonlinear Model Predictive Control (NMPC) onto the Unitree Go2 quadruped robot. The authors implement and validate real-time NMPC for simultaneous locomotion and object interaction tasks directly on the Go2 platform, leveraging its onboard computing and actuation capabilities. The work demonstrates practical whole-body control on a real Unitree robot, offering insights into embedded optimization-based control for agile quadrupeds.

• "Transformable Quadruped Wheelchairs Capable of Autonomous Stair Ascent and Descent" (2024) PDF (Citations: 1) — Sensors — This paper presents a transformable quadruped wheelchair inspired by the Unitree B2 robot, integrating wheels for flat terrain and robotic legs for stair navigation. It uses reinforcement learning with curriculum learning in a Unity-based simulation to train stair ascent and descent behaviors, achieving high success rates. While the design demonstrates promising mobility assistance potential, it currently struggles with complex stair types and descent stability, highlighting the need for further safety enhancements.

• "A Tracking Control Approach With Sequence-Scaling Lyapunov-Based MPC for Quadruped Robots" (2024) (Citations: 7) — IEEE Transactions on Industrial Informatics — This paper proposes a sequence-scaling Lyapunov-based model predictive control (MPC) algorithm to enhance tracking performance for autonomous quadruped robots. The method introduces a unified kinematic model for translation and rotation modes and ensures closed-loop stability via a contraction constraint embedded in the gait sequence, while respecting dynamic balance and speed limits. The approach is validated through both simulation and real-world experiments on the Unitree Aliengo, demonstrating strong real-time control capabilities.

• "Sailing Through Point Clouds: Safe Navigation Using Point Cloud Based Control Barrier Functions" (2024) PDF — arXiv (Cornell University) — This paper introduces a novel control barrier function (CBF)-based local planner called 'Sailing Through Point Clouds,' featuring two components: Vessel, a scaling factor-based CBF that operates directly on point cloud data, and Mariner, a preview control framework to avoid spurious equilibria. The method is experimentally validated on Unitree B1 and Unitree Go2 quadruped robots across diverse real-world environments, demonstrating safe navigation without relying on pre-built maps. The work provides a practical Sim2Real-ready safety layer compatible with global planners, enhancing autonomous locomotion for Unitree legged robots.

• "A Study of Climbing Strategy on Unitree Aliengo" (2023) — This paper investigates climbing strategies specifically designed for the Unitree Aliengo quadruped robot, focusing on gait adaptation and terrain negotiation. The authors develop and test control policies that enable Aliengo to traverse inclined surfaces using dynamic locomotion patterns, validated through real-world experiments on the physical robot. The work contributes practical insights into legged locomotion on challenging terrains using Unitree's hardware platform.

• "Hierarchical Optimization-based Control for Whole-body Loco-manipulation of Heavy Objects" (2023) PDF — arXiv (Cornell University) — This paper introduces a hierarchical optimization-based control framework for whole-body loco-manipulation of heavy objects using legged robots. It integrates an online manipulation planner, pose optimization for kinematic alignment, and a linear MPC controller that accounts for manipulation forces. The method is experimentally validated on a Unitree Aliengo robot fitted with a custom robotic arm, demonstrating successful lifting and transport of an 8kg payload and door manipulation, highlighting the importance of whole-body coordination over standard locomotion MPC.

• "Dynamic Hybrid Locomotion and Jumping for Wheeled-Legged Quadrupeds" (2023) (Citations: 3) — This paper presents a motion optimization framework enabling wheeled-legged quadrupeds to perform dynamic hybrid locomotion and jumping over obstacles without reducing speed. It employs a model predictive controller based on a time-varying rigid body dynamics model that accounts for both legs and non-steerable wheels, and introduces an energy-efficient driving strategy with minimal leg swings. The method is experimentally validated on both the Mini Cheetah and the Unitree Aliengo, demonstrating successful obstacle traversal via dynamic jumps. This work provides a practical approach to enhancing mobility and efficiency in real-world wheeled-legged robot navigation.

🔧 Projects

• go2_omniverse — This project provides Unitree Go2 and G1 robot support for NVIDIA Isaac Lab, enabling simulation and reinforcement learning workflows in Isaac Sim/Omniverse. It includes robot URDFs, sensor configurations, and integration with the Orbit framework for Sim2Real transfer and robot learning experiments. The repository targets researchers and developers working on legged locomotion and whole-body control using Unitree's Go2 and G1 platforms.

• unitree_rl_lab — This repository provides reinforcement learning implementations specifically designed for Unitree robots using the IsaacLab simulation framework. It includes training environments, reward functions, and policy architectures tailored to Unitree platforms like Go2 and H1, enabling rapid Sim2Real transfer. The project is aimed at researchers and developers working on legged locomotion and whole-body control for Unitree robots.

• autonomy_stack_go2 — This project provides a full autonomy stack specifically designed for the Unitree Go2 quadruped robot, enabling autonomous navigation and perception in real-world environments. Built in C++, it integrates modules for state estimation, mapping, planning, and control tailored to Go2's hardware and sensor suite. The stack is aimed at researchers and developers working on legged robot autonomy who need a robust, integrated baseline for deployment.

• go2_robot — This repository provides official ROS 2 packages and hardware interfaces for the Unitree Go2 quadruped robot, enabling low-level control, sensor integration, and real-time communication with the robot's onboard systems. It includes drivers for joint control, IMU, and depth cameras, and is designed for developers building autonomous navigation or locomotion applications on the Go2 platform.

• unitree_go2_ros2 — This project provides a complete ROS 2 Jazzy integration for the Unitree Go2 quadrupedal robot, leveraging the CHAMP controller framework to enable real-time control and state estimation. It includes hardware interfaces, sensor drivers, and locomotion controllers specifically tailored for the Go2 platform. The package is designed for robotics researchers and developers working with Unitree Go2 in ROS 2 environments.

• Awesome-Unitree-Projects — This repository curates a list of awesome projects built around Unitree robots, serving as a community-driven index for developers and researchers. It aggregates external resources such as codebases, tutorials, and research implementations that target specific Unitree platforms like Go2, H1, or G1. The project itself does not provide original tools or integrations but acts as a meta-list pointing to other Unitree-related work.

• go2_ros2_sdk — This project provides an unofficial ROS 2 SDK for the Unitree Go2 series (AIR/PRO/EDU), enabling developers to interface with the robot's low-level control, state feedback, and sensor data through ROS 2 topics and services. It includes Python-based drivers and message definitions tailored specifically for Go2 hardware, facilitating integration into ROS 2-based robotic applications. The repository is actively maintained and widely used by the community for research and development on Unitree Go2 platforms.

• FAST_LIO_LOCALIZATION_HUMANOID — This project provides a LiDAR-based localization solution specifically designed for humanoid robots such as the Unitree G1. Built on the FAST-LIO framework, it enables real-time, tightly coupled lidar-inertial odometry optimized for humanoids with dynamic motion patterns. The system integrates IMU and LiDAR data using an error-state Kalman filter and supports deployment on actual hardware like the G1, making it valuable for researchers and developers working on autonomous navigation for Unitree humanoids.

• walk-these-ways-go2 — This project adapts the 'Walk These Ways' reinforcement learning framework for the Unitree Go2 quadruped robot, enabling agile locomotion policies trained in Isaac Gym and transferred to real hardware via Sim2Real. It includes C++ deployment code tailored to Go2's SDK and motor control interface, allowing researchers and developers to reproduce and extend high-performance walking behaviors on the actual robot.

• LocomotionWithNP3O — This project implements locomotion learning for the Unitree Go2 robot using the N-P3O reinforcement learning algorithm and a HIM-like policy, trained in the Isaac Gym simulation environment. It provides a complete pipeline from simulation training to real-world deployment on the Go2 platform, featuring domain randomization and sim-to-real transfer techniques. The repository is aimed at researchers and developers working on quadrupedal robot locomotion with Unitree hardware.

• isaac-go2-ros2 — This project provides a Unitree Go2 simulation platform built on NVIDIA Isaac Sim and ROS 2 Humble, enabling development and testing of autonomous navigation, decision-making, and perception tasks. It integrates camera and LiDAR sensors with Isaac ROS components and supports deployment via Isaac Lab for reinforcement learning or control algorithms. The repository is tailored for robotics researchers and developers working specifically with the Unitree Go2 quadruped.

• humanoid_amp — This project implements Humanoid AMP (Adversarial Motion Priors) in Isaac Lab specifically for the Unitree G1 humanoid robot. It enables realistic motion imitation and policy learning by leveraging adversarial training with motion capture data, targeting researchers and developers working on humanoid locomotion and control. The repository provides Isaac Lab integration, G1-specific asset configurations, and training scripts tailored for Unitree's hardware.

• dddmr_navigation — dddmr_navigation is a comprehensive 3D navigation stack for mobile robots that includes modules for mapping, localization, perception, path planning, and control. It explicitly supports the Unitree Go2 quadruped robot, integrating with its hardware for real-world deployment, and leverages 3D SLAM and perception pipelines tailored for legged platforms. The system is implemented in C++ and targets researchers and developers working on autonomous navigation for quadrupedal robots like the Unitree Go2.

• unitree_webrtc_connect — This project provides a WebRTC driver for Unitree Go2 and G1 robots, enabling real-time remote control and video streaming through WebRTC protocols. It includes Python-based interfaces for low-latency communication and supports both robot models with hardware-specific integration. The tool is aimed at developers and researchers needing browser-based teleoperation or remote monitoring capabilities for Unitree robots.

• unitree-go2-ros2 — This project provides a ROS 2 integration for the Unitree Go2 quadruped robot, leveraging the CHAMP legged robotics framework to enable simulation in Gazebo and hardware control. It includes URDF models, ROS 2 control configurations, and launch files tailored specifically for the Go2 platform. The repository targets researchers and developers working on legged locomotion with Unitree robots in ROS 2 environments.

• g1_spinkick_example — This project demonstrates training a Unitree G1 humanoid robot to perform a double spin kick using the MuJoCo-based MJLab framework. It leverages mujoco-warp for physics simulation and includes Sim2Real transfer techniques to deploy the learned policy on the physical G1 robot. The implementation provides a complete pipeline from motion design to real-world execution, targeting researchers and developers working on dynamic humanoid locomotion and martial arts-inspired behaviors.

• GR00T-WholeBodyControl — This project provides a software stack for loco-manipulation experiments with strong support for the Unitree G1 humanoid robot. It includes whole-body control policies, a teleoperation system, and tools for data export, enabling advanced research in dynamic locomotion and manipulation. Built primarily in Python, it targets researchers and developers working on humanoid robotics using the Unitree G1 platform.

• OpenWBC — OpenWBC is a VR-based teleoperation and data collection system designed specifically for the Unitree G1 humanoid robot, enabling whole-body control through immersive virtual reality interfaces. It integrates low-latency motion capture with real-time robot command execution, supporting vision-language-action (VLA) learning pipelines. The system provides tools for collecting high-quality human demonstration datasets, targeting researchers developing embodied AI and humanoid control algorithms.

• go2-webrtc — This project provides a WebRTC API for Unitree GO2 robots, enabling real-time communication and control over the robot via WebRTC protocols. It includes Python-based implementation for streaming sensor data and commanding robot actions through a web interface, making it suitable for remote teleoperation and monitoring. The tool is tailored specifically for GO2 users seeking low-latency, browser-accessible robot interaction.

• g1pilot — hucebot/g1pilot is a ROS2 package specifically designed for the Unitree G1 humanoid robot, providing core functionalities such as inverse kinematics, teleoperation, and navigation. Built in Python, it enables developers to interface with the G1's hardware through ROS2 topics and services, facilitating tasks like whole-body motion control and remote operation. The project targets robotics researchers and developers working on humanoid locomotion and autonomy using the Unitree G1 platform.

• unitree-go2-slam-nav2 — This project integrates SLAM and autonomous navigation capabilities specifically for the Unitree-Go2 robot using ROS 2 and Nav2. It provides a complete pipeline for real-time mapping and path planning by leveraging the Go2's onboard sensors and locomotion system. The implementation is tailored to the Unitree-Go2's hardware interface, making it a practical solution for developers working on autonomous quadruped navigation.

• unitree_rl_mjlab — This repository provides reinforcement learning implementations specifically designed for Unitree robots using the MuJoCo physics simulator. It includes training environments and example policies tailored for Unitree platforms, enabling researchers and developers to prototype locomotion and control strategies in simulation before real-world deployment.

• go2_isaac_ros2 — This project provides a simulation environment for the Unitree Go2 robot in NVIDIA Isaac Sim, leveraging ROS2 for low-level joint control. It enables developers to test and develop control algorithms in a high-fidelity simulated setting before deploying to real hardware. The integration with ROS2 facilitates modular robotics development and is particularly useful for researchers and engineers working on legged locomotion.

• qm_door — This project implements a quadruped manipulator planner and controller for the Unitree Aliengo equipped with a Z1 robotic arm, focusing on complex tasks like door opening and stair navigation. It combines nonlinear model predictive control (MPC) and whole-body control (WBC) using the OCS2 framework, and integrates YOLOv8n for perception. The system is designed for researchers and developers working on mobile manipulation with Unitree's quadruped platforms.

• go2-convex-mpc — This project implements a convex Model Predictive Control (MPC) locomotion controller specifically for the Unitree Go2 quadruped robot, using MuJoCo for simulation and Pinocchio for kinematic and dynamic computations. It provides a real-time capable, optimization-based gait controller tailored to the Go2's hardware constraints and dynamics, enabling stable walking and trotting gaits. The repository is aimed at robotics researchers and developers working on advanced legged locomotion for Unitree platforms.

• go2_parkour_deploy — This project enables deployment of parkour locomotion policies trained in IsaacLab to both MuJoCo simulations and real-world Unitree Go2 robots. It provides Sim2Sim and Sim2Real transfer pipelines, leveraging Python-based control interfaces and environment wrappers for the Go2 platform. The repository is aimed at researchers and developers working on agile legged locomotion and sim-to-real transfer for quadrupedal robots.

• Go2-Dynamic-Inspection — This project provides an unofficial ROS2 SDK for the Unitree Go2 series (AIR/PRO/EDU) with integrated 3D LiDAR support, enabling autonomous navigation and inspection tasks. It combines DLIO for robust odometry, FAR Planner for local path planning, and Open3D-SLAM for real-time mapping, all within a ROS 2 Humble framework. The repository is tailored for robotics researchers and developers aiming to extend the Go2's perception and autonomy capabilities using C++.

• G1-retarget — This project uses the Physics-based Humanoid Control (PHC) framework to retarget human motion data from the AMASS dataset into motions suitable for the Unitree G1 humanoid robot. It provides a pipeline for adapting complex human movements to the G1's kinematic and dynamic constraints, leveraging simulation for validation. The tool is aimed at researchers and developers working on humanoid motion generation and Sim2Real transfer for the Unitree G1 platform.

• phase_guided_terrain_traversal — This project presents a perceptive reinforcement learning framework for terrain traversal specifically designed for the Unitree Go2 robot. It uses MuJoCo Playground for simulation training and deploys policies on real hardware via unitree_sdk2py, enabling adaptive locomotion over rough terrain. The implementation leverages phase-guided control and sensory feedback, targeting researchers and developers working on legged locomotion with Unitree robots.

• My_unitree_go2_gym — This project provides a custom Gym environment for reinforcement learning with the Unitree Go2 quadruped robot. It includes Python-based interfaces for simulation and control, likely built on common RL frameworks, enabling researchers and developers to train and test locomotion policies specifically tailored for the Go2 platform.

• go2_firmware_tools — This project provides Python-based firmware tools specifically designed for the Unitree Go2 robot, enabling users to interact with and manage low-level firmware functionalities. It includes utilities for firmware flashing, parameter configuration, and diagnostics, directly targeting Go2 hardware interfaces. The repository is actively maintained and serves developers and researchers working on customizing or troubleshooting Unitree Go2 firmware.

• go2_ros2_toolbox — This project provides a ROS 2 toolbox specifically designed for the Unitree Go2 robot, focusing on SLAM and navigation capabilities. It includes C++ implementations leveraging ROS 2 interfaces to enable autonomous mapping and path planning on the Go2 platform. The toolbox is aimed at robotics developers and researchers working with Unitree Go2 in real-world or simulated environments.

• robot-unitree-g1 — This project focuses on research and learning for the Unitree G1 humanoid robot using the MuJoCo simulation environment. It provides Python-based implementations for simulating and experimenting with G1's manipulation capabilities, leveraging MuJoCo's physics engine for realistic dynamics. The repository is aimed at researchers and developers interested in humanoid robot control and manipulation tasks.

• unitree_go2_nav — This project provides a navigation and SLAM solution specifically designed for the Unitree Go2 robot, leveraging RTAB-Map and custom high-level control built on the Unitree ROS2 SDK. It enables autonomous navigation in real-world environments using ROS2, with integration tailored to the Go2's hardware interface. The package is aimed at robotics researchers and developers working on legged robot autonomy.

• lecabot — LeCabot is a low-cost mobile manipulator modification designed specifically for the Unitree Go2 and SO-100 platforms, enabling robotic arm integration for enhanced manipulation capabilities. The project provides hardware designs and control interfaces tailored to the Go2's onboard systems, targeting researchers and hobbyists interested in affordable mobile manipulation. It directly extends the functionality of the Unitree Go2, making it relevant to the Unitree robotics ecosystem.

• unitree-go2-mcp-server — The Unitree Go2 MCP Server enables natural language control of the Unitree Go2 robot by integrating a Large Language Model (LLM) via the Model Context Protocol (MCP). Built in Python and leveraging ROS2, it translates high-level commands into actionable robot behaviors, offering an accessible interface for non-expert users. This project is tailored specifically for the Unitree Go2 and demonstrates a practical application of LLMs in quadrupedal robot teleoperation.

• G1_deploy — This project focuses on deploying reinforcement learning policies directly onto the Unitree G1 humanoid robot. It provides Python-based tools and scripts to interface with the G1's control system, enabling real-world execution of trained policies. The repository includes utilities for hardware communication, state estimation, and low-level actuation, targeting researchers and developers working on real-world humanoid locomotion and control.

• Go2Py — Go2Py provides a Python interface and simulation environment specifically designed for the Unitree Go2 quadruped robot. It enables high-level control, state estimation, and policy deployment using a C++ backend with Python bindings, supporting both real hardware and simulated environments. The project is tailored for researchers and developers working on locomotion, reinforcement learning, or whole-body control for the Go2 platform.

• go2_python_sdk — This project provides an unofficial Python SDK for the Unitree Go2 robot using DDS (Data Distribution Service) for communication. It enables developers to interface with the Go2's low-level control, state feedback, and sensor data through a Python API, leveraging the robot's native DDS middleware. The SDK is aimed at researchers and developers seeking flexible, high-level control of the Unitree Go2 without relying on official C++ tooling.

• unitree_cpp — This project enables wireless control of the Unitree G1 humanoid robot by providing a C++ interface that eliminates the need for cable connections. It leverages pybind11 to expose low-level robot functionalities to Python, facilitating real-time command and sensor data exchange over Wi-Fi. The repository includes utilities for UDP communication and hardware abstraction, targeting developers building teleoperation or autonomous applications for the Unitree G1.

• RoboMimicDeploy_G1 — RoboMimicDeploy_G1 is a policy deployment and control framework specifically designed for the Unitree G1 robot, enabling seamless transition between Mujoco-based simulation and real-world execution. It provides real-time control capabilities and integrates with the G1's hardware interface for deploying learned locomotion or manipulation policies. The project targets researchers and developers working on sim-to-real transfer and embodied AI for humanoid robots.

• amigo_ros2 — This project provides ROS 2 support for the Unitree Go2 quadruped robot, enabling integration with the Robot Operating System 2 ecosystem. It includes C-based drivers and interfaces to access low-level control, sensor data, and state estimation from the Go2 platform. The repository is actively maintained and targets developers seeking to build autonomous applications using standard ROS 2 tools on the Unitree Go2.

• Unitree_GO2_SBUS — This project provides C++ code for interfacing with the Unitree Go2 robot using SBUS protocol, enabling remote control and telemetry communication. It includes low-level drivers and example implementations for SBUS-based command handling, specifically tailored for the Go2 platform. The repository targets developers working on custom remote control solutions or integrating third-party controllers with Unitree Go2 robots.

• InternNav-deploy — This project provides an edge deployment guide for running InternNav-based perception and navigation systems on Unitree Go2, Go2W, and B2 robots using ROS 2, RealSense cameras, and Python. It focuses on real-world deployment of vision-based navigation, offering hardware integration instructions and optimized inference pipelines tailored for Unitree quadrupeds. The repository is aimed at robotics developers seeking to deploy autonomous navigation on resource-constrained Unitree platforms.

• unitree_go2_ros — This project provides ROS 1 drivers for the Unitree Go2 robot using a Python-based WebRTC interface, enabling real-time communication and control. It leverages the robot's native WebRTC API to publish sensor data (e.g., IMU, joint states) and accept velocity commands, offering a lightweight alternative to official SDKs. Targeted at ROS developers seeking rapid integration of the Go2 into existing robotics workflows.

• unitree-go2-slam-toolbox — This project provides a ROS2-based SLAM toolbox tailored for the Unitree Go2 quadruped robot, integrating visual robot modeling, PointCloud2-to-Laserscan conversion, and compatibility with standard SLAM algorithms. It enables real-time 2D mapping using the Go2’s onboard sensors by bridging 3D point cloud data to 2D laser scan formats required by navigation stacks. Designed for robotics developers working with Unitree Go2 in ROS2 environments, it facilitates rapid deployment of autonomous navigation capabilities.

• Multi-Hetero-Agent-Exploration-on-UnitreeGOs — This project enables heterogeneous collaborative exploration using a fleet of Unitree GO1 and GO2 quadrupeds, integrating ROS2-based navigation, SLAM, and multi-agent coordination in C++. It provides robot-specific configurations and communication frameworks tailored for Unitree Go series hardware, targeting researchers and developers working on multi-robot autonomous exploration.

• unitree_guide_go2 — 该项目将Unitree_Guide框架适配到Unitree Go2机器人实机，提供C++实现的底层控制接口和运动控制示例。它支持Go2的实时通信与状态反馈，便于开发者快速部署自定义控制算法。主要面向希望在Go2平台上进行运动控制、步态开发或算法验证的机器人研究人员和工程师。

• unitree_go2w_agent_sdk — This project provides a unified C-based SDK for the Unitree Go2W quadruped, offering a single API to integrate perception, planning, and control modules tailored for AI agent development. It enables seamless low-level hardware interaction and high-level autonomy on the Go2W platform, targeting robotics researchers and developers building intelligent agents for Unitree robots.

• unitree-g1-autonomous — This project implements a fully autonomous navigation system for the Unitree G1 humanoid robot by leveraging AI-powered visual analysis through the Google Gemini API. It integrates computer vision and large language model capabilities to enable environment perception and decision-making, with Python-based control logic interfacing directly with the G1's hardware. The system is designed for researchers and developers aiming to explore high-level autonomous behaviors on the Unitree G1 platform.

• RCI_quadruped_robot_navigation — This project integrates reinforcement learning, navigation stacks, and Open-RMF to enable autonomous driving capabilities specifically for the Unitree Go2 and Go2W quadruped robots. It provides a C++-based framework that bridges high-level task planning with low-level locomotion control on real hardware. The repository targets researchers and developers working on autonomous navigation for Unitree’s Go2 series.

• Deploy-an-RL-policy-on-the-Unitree-Go2-robot — This project provides a ROS 2-based framework for deploying reinforcement learning (RL) policies on the Unitree Go2 quadruped robot, using MuJoCo for simulation-to-real (Sim2Real) transfer. By toggling the is_simulation ROS parameter, users can seamlessly switch between simulated validation and real-world execution. It targets researchers and developers working on legged locomotion with RL on Unitree hardware.

• go2_odometry — This project provides simple state estimation specifically designed for the Unitree Go2 robot, implementing odometry algorithms in Python to estimate the robot's pose and motion. It directly targets Unitree Go2 hardware, leveraging its sensor data for real-time state tracking. The repository is maintained by the INRIA Paris Robotics Lab and serves researchers and developers working on Go2 navigation and control.

• G1_AMO_control — This project enables VR-based teleoperation for the Unitree G1 humanoid robot, allowing users to control the robot's movements through virtual reality input. It integrates with the G1's API to map VR controller poses to whole-body motion commands, facilitating intuitive remote operation. The system is designed for researchers and developers exploring immersive telepresence and human-robot interaction with the Unitree G1 platform.

• IK-humanoid — This project implements inverse kinematics (IK) to enable partial hand movements on the Unitree G1 humanoid robot, leveraging the unitree-rl-gym framework and IsaacGym for simulation. It provides a Python-based solution for controlling the G1's upper limbs with IK solvers, targeting researchers and developers working on dexterous manipulation for Unitree's humanoid platform.

• g1-isaac-groot-n1 — This project focuses on fine-tuning and deploying NVIDIA's Isaac GR00T N1 foundation model specifically for the Unitree G1 humanoid robot. It integrates the GR00T N1 policy with the G1's hardware interface using Python, leveraging NVIDIA Isaac Sim for simulation and aiming for real-world deployment. The work targets researchers and developers working on vision-language-action models for humanoid robots, particularly those using the Unitree G1 platform.

• Go2_Isaac_ros2 — This project provides a simulation environment for the Unitree Go2 robot using NVIDIA IsaacLab/Isaac-Sim integrated with ROS2, enabling navigation and other robotic tasks. It leverages Isaac Sim's physics engine and ROS2 middleware to facilitate development and testing of autonomous behaviors in a simulated setting. The repository is aimed at researchers and developers working on legged locomotion and mobile manipulation with the Unitree Go2 platform.

• himloco_lab — This project enables training, exporting, and deploying HimLoco locomotion policies specifically for the Unitree Go2 robot within the Isaac Lab simulation environment. It leverages reinforcement learning to develop agile and robust control policies, with direct integration for the Go2's hardware interface and dynamics. The repository is aimed at researchers and developers working on legged locomotion and Sim2Real transfer for Unitree robots.

• unitree-go2-mjx-rl — This project implements reinforcement learning for quadruped locomotion on the Unitree Go2 robot using Mujoco XLA (MJX), a high-performance physics simulation framework optimized for JAX. It provides a lightweight, differentiable environment tailored specifically for training and simulating Go2 policies with accelerated computation on GPUs/TPUs. The repository includes reward shaping, gait control strategies, and Sim2Real transfer considerations, targeting researchers and developers working on agile legged locomotion for Unitree robots.

• unitree_g1_python — This project provides a Python adaptation of Unitree's SDK2 specifically tailored for the G1 humanoid robot, including simplified interfaces and practical examples. It modifies the original SDK to focus exclusively on G1 functionality, offering developers easier access to control and interaction capabilities. The repository is aimed at researchers and engineers working with the Unitree G1 platform.

• Unitree_G1_pose_control-through-UI — This project provides a motion-design assistant tool specifically for the Unitree G1 humanoid robot, allowing users to intuitively create and fine-tune poses in a 3D interface and transmit them to the robot's control system for accurate real-world execution. Built in Python, it bridges graphical pose design with physical robot actuation, targeting developers and researchers working on humanoid motion planning and teleoperation.

• unitree_h1_carrybox — This project provides a ROS-based framework for dexterous manipulation tasks using the Unitree H1 humanoid robot, specifically focusing on box-carrying scenarios. It integrates the H1's whole-body control with its dexterous hands through C++ modules and ROS communication, enabling coordinated arm and hand movements. The repository targets researchers and developers working on humanoid manipulation with the Unitree H1 platform.

• go2_navigation — This project provides a complete autonomous navigation pipeline for the Unitree Go2 robot, integrating the Livox MID360 LiDAR for perception. It implements SLAM and path planning capabilities specifically tailored for the Go2's hardware and mobility constraints. The system is built in C++ and targets researchers and developers working on autonomous navigation with Unitree quadrupeds.

• mjlab-homierl — This project is a fork of mjlab that implements the HOMIE locomotion policy training framework from an RSS 2025 paper, with explicit support for the Unitree H1 humanoid robot and Robotiq 2F85 gripper. It uses MuJoCo for simulation and focuses on human-in-the-loop teleoperation and loco-manipulation learning. The repository targets researchers working on humanoid robotics and Sim2Real transfer for Unitree H1 platforms.

• Go2-EDU-Mycobot-320-M5-Gazebo-Simulation-with-Navigation-and-MoveIt — This project provides a Gazebo-based ROS simulation integrating the Unitree Go2 EDU quadruped with the Mycobot 320 M5 arm, featuring autonomous navigation via the ROS Navigation Stack for the Go2 and motion planning using MoveIt! for the arm. It enables coordinated mobile manipulation tasks in simulation, targeting researchers and developers working on heterogeneous robot systems involving Unitree's Go2 platform.

• unitree-g1-bipedal-rl-walk — This project implements reinforcement learning for bipedal walking on the Unitree G1 robot using Isaac Gym for training and MuJoCo for evaluation. It provides a complete RL pipeline tailored specifically to the G1's kinematics and dynamics, enabling sim-to-real transfer research. The code is aimed at researchers and developers working on legged locomotion for Unitree's humanoid platform.

• unitree_go2w_ros2 — This project enables users to create a digital twin of the Unitree Go2W quadruped robot in Gazebo using ROS2, providing a simulated environment for development and testing. It includes URDF models, ROS2 control interfaces, and Gazebo plugins tailored specifically for the Go2W variant, facilitating realistic simulation of its dynamics and sensors. The repository is aimed at robotics developers and researchers working with Unitree's Go2 platform who need a reliable simulation setup for algorithm prototyping.

• unitree_g1_optimal_controllers — This project implements an SRBM-MPC (Single Rigid Body Model Model Predictive Control) controller specifically designed for the Unitree G1 humanoid robot, using Simulink Simscape Multibody for simulation. It provides a MATLAB-based framework for testing and developing optimal control strategies on the G1 model, leveraging Simscape Multibody's physics engine for accurate dynamics. The repository is aimed at researchers and engineers working on advanced motion planning and whole-body control for the Unitree G1 platform.

• go2_pressure_sensor — This project provides an add-on pressure sensor module specifically designed for the Unitree Go2 robot to enable accurate foot contact detection. It includes hardware integration instructions and software interfaces compatible with the Go2's control system, allowing developers to enhance locomotion algorithms with real-time ground contact feedback. The solution is targeted at robotics researchers and engineers working on legged locomotion and terrain adaptation for the Unitree Go2 platform.

• elder_and_dog — This project builds a home companion and object-finding system using the Unitree Go2 quadruped robot, integrating an MCP + LLM Agent architecture to interpret natural language commands like 'find my water.' It leverages Vision-Language Models (VLM) for environmental understanding and autonomous navigation with obstacle avoidance. Designed specifically for elderly users, it aims to address loneliness and misplaced items by making the Go2 an intuitive, responsive companion.

• unitree_mpc — This project implements a Model Predictive Control (MPC) framework for the Unitree G1 humanoid robot using the Optimal Control Software Suite (OCS2). It provides whole-body motion planning and control capabilities specifically tailored for the G1's kinematic and dynamic constraints, leveraging OCS2's advanced optimization tools. The implementation is designed for researchers and developers working on advanced locomotion and manipulation tasks with the Unitree G1 platform.

• g1_locomotion — This project implements a humanoid locomotion stack specifically for the Unitree G1 robot using convex Model Predictive Control (MPC). It provides a Python-based control framework that enables dynamic walking and balance maintenance on the G1 platform, leveraging real-time optimization for gait generation and stability. The repository targets researchers and developers working on advanced bipedal locomotion for the Unitree G1.

• G1_localization — This project provides a localization toolbox specifically designed for the Unitree G1 humanoid robot, implementing C++-based algorithms to enable accurate pose estimation in indoor environments. It integrates with the G1's sensor suite and likely leverages ROS or similar middleware for real-time operation, targeting researchers and developers working on autonomous navigation for Unitree's humanoid platform.

• ros2_go2_video — This project provides a ROS 2 node written in Rust that publishes the camera or video stream from the Unitree Go2 robot into the ROS 2 ecosystem. It enables real-time access to Go2's visual data for perception, navigation, or teleoperation tasks by bridging the robot's native video output with standard ROS 2 image topics. The tool is particularly useful for developers integrating Go2 into ROS 2-based robotic applications.

• auki_robotics_g1_humanoid_ros2 — This project provides ROS2 packages specifically designed for the Unitree G1 humanoid robot, enabling integration with the Robot Operating System 2 framework. It includes drivers, interfaces, and utilities tailored to the G1's hardware and control architecture, facilitating tasks like motion control, sensor data handling, and system monitoring. The repository is aimed at developers and researchers working on humanoid robotics applications using the Unitree G1 platform.

• unitree-go2-realtime-agent — This project implements an AI agent for the Unitree Go2 EDU quadruped robot that bridges the OpenAI Realtime API with WSO2 middleware and the official Unitree Go2 SDK. It enables real-time voice interaction and command execution on the robot using Python, leveraging WSO2's IoT integration capabilities. The tool is aimed at developers exploring conversational AI applications on Unitree’s Go2 platform.

• Aligator_Unitree_G1 — This project models the walking behavior of the Unitree G1 robot using the Aligator trajectory optimization library. It provides a Python-based implementation for generating and optimizing locomotion trajectories specifically tailored for the G1's kinematic and dynamic constraints. The work is relevant for researchers and developers focused on advanced motion planning and control for Unitree's humanoid platform.

• humanoid-motion-planning — This project provides whole-body motion planning for the Unitree G1 humanoid robot in MuJoCo, integrating ZMP preview control, A* footstep planning, MPC-based balance control (achieving 49% energy reduction), reinforcement learning for locomotion, and Jacobian-based inverse kinematics for manipulation. It offers a comprehensive Python framework targeting researchers and developers working on advanced humanoid locomotion and control specifically for the Unitree G1 platform.

• elevation_mapping_g1 — This project implements GPU-accelerated elevation mapping specifically for the Unitree G1 humanoid robot. It processes sensor data to build real-time 3D elevation maps, leveraging C++ for performance-critical components and integrating with the G1's perception pipeline. The tool is aimed at developers and researchers working on terrain-aware navigation for the Unitree G1.

• Cerebro-Control — Cerebro-Control provides a VR-based teleoperation system specifically designed for the Unitree H1 humanoid robot, leveraging its built-in sports mode that utilizes reinforcement learning for locomotion. The project integrates VR input with the robot's native control stack to enable intuitive remote operation while relying on the H1's onboard RL policies for stable movement. This setup is aimed at developers and researchers exploring immersive telepresence applications with Unitree's advanced humanoid platform.

• Quadrupeds_Climbing — This project implements a reinforcement learning task in Isaac Lab to train a Unitree Go2 quadruped robot for climbing steep terrain. It leverages Isaac Lab's simulation environment to develop and test locomotion policies specifically tailored for the Go2's dynamics and kinematics. The repository provides a complete training setup, making it useful for researchers and developers working on legged locomotion and Sim2Real transfer for Unitree robots.

• IsaacLab_Locomotion_H1 — This project extends IsaacLab to enable rough terrain locomotion for the Unitree H1 humanoid robot. It provides reinforcement learning environments and training frameworks specifically tailored for H1's dynamics and actuation, leveraging IsaacLab's GPU-accelerated simulation. The repository includes reward formulations, observation spaces, and policy architectures designed for challenging terrains, targeting researchers and developers working on humanoid locomotion.

• IsaacLab_Locomotion_H1-2 — This project provides an IsaacLab-based extension for locomotion control of the Unitree H1-2 humanoid robot on both flat and rough terrains. It implements reinforcement learning policies tailored to the H1-2's kinematics and dynamics, leveraging IsaacLab's high-fidelity simulation environment for training and evaluation. The repository is aimed at researchers and developers working on humanoid locomotion and Sim2Real transfer for Unitree robots.

• g1_teleop — This project enables teleoperation of the Unitree G1 humanoid robot using Nintendo Joy-Cons for arm control while retaining locomotion via the official remote. It implements a hybrid control mode and supports both real hardware and MuJoCo simulation, leveraging Python for interface and control logic. Targeted at developers and researchers working on intuitive teleoperation systems for the G1 platform.

• unitree_go2_deploy — This project provides sim2sim (MuJoCo) and sim2real deployment capabilities specifically for the Unitree Go2 quadruped robot. It includes tools and workflows to transfer locomotion policies trained in simulation to the real-world Unitree Go2 platform, leveraging MuJoCo for high-fidelity simulation. The repository is aimed at researchers and developers working on legged locomotion and Sim2Real transfer for Unitree robots.

• Go2_planner_suite — This project provides a complete autonomous navigation stack tailored for the Unitree Go2 quadruped robot, integrating components like a far-planner and Fast-LIO2 for robust localization and path planning. Built on ROS 2 Humble and written in C++, it enables real-time perception, mapping, and navigation in dynamic environments. The suite is designed for robotics researchers and developers aiming to deploy autonomous capabilities on the Go2 platform.

• go2-rl-locomotion — This project provides a reinforcement learning framework specifically designed for quadruped locomotion on the Unitree Go2 robot. It supports multiple RL algorithms (PPO, SAC, TD3, DDPG), incorporates advanced techniques like reward shaping, curriculum learning, and domain randomization, and includes visualization tools for training analysis. Built in Python, it targets researchers and developers working on agile and robust gait control for the Go2 platform.

• unitree_walking — This project implements walking control for the Unitree G1 robot using PPO reinforcement learning trained in Isaac Lab and simulated in NVIDIA Omniverse. It provides a complete RL pipeline including environment setup, reward design, and policy deployment specifically tailored for the G1's kinematics and dynamics. The repository targets researchers and developers working on bipedal locomotion and Sim2Real transfer for Unitree humanoid platforms.

• legged_lab — This project leverages IsaacLab to train legged robots with a focus on Unitree platforms, implementing fall recovery and blind walking for the Unitree A1, continuous backflips on the same model, and dance imitation capabilities for the Unitree H1. It provides IsaacLab-based reinforcement learning environments and motion control policies tailored to these specific Unitree robots. The repository targets researchers and developers working on agile locomotion and whole-body control for Unitree’s quadruped and humanoid systems.

• go2-autonomous-patrol — This project implements an autonomous security and monitoring system specifically for the Unitree Go2 quadruped robot. It leverages ROS 2 and Nav2 for navigation and path planning, and includes a web-based dashboard for remote control and monitoring. The system is designed for developers and security robotics teams seeking to deploy autonomous patrol capabilities on Unitree's Go2 platform.

• Unitree-G1-Sim2Real — This project provides a minimal Sim2Real template specifically designed for deploying reinforcement learning policies on the Unitree G1 humanoid robot. It includes C++ implementations that bridge simulation-trained policies to real-world execution on the G1 platform, leveraging the robot's native control interfaces. The repository serves as a lightweight starting point for researchers and developers working on RL-based locomotion or manipulation tasks with the Unitree G1.

• JeffrinSam_MTS — This project generates synthetic joint data for Unitree H1 and G1 robots using imitation learning within Isaac Lab. It leverages Isaac Sim for simulation and visualization, enhances scene realism with NVIDIA Cosmos, and integrates teleoperation data to fine-tune the Gr00t N1.5 model for pick-and-place tasks—all validated through inference in Isaac Sim. The work targets researchers and developers focused on Sim2Real transfer and dexterous manipulation for Unitree humanoid platforms.

• MPC-h1-Manipulation-task — This project implements a two-layer Model Predictive Control (MPC) system specifically for arm manipulation tasks on the Unitree H1 robot, using a reduced 8-DOF arm model. It features a high-level Trajectory MPC for end-effector path planning and a low-level Kinematics MPC for joint motion execution, targeting researchers and developers working on advanced manipulation with the H1 platform.

• unitree_go2w_ros2 — This project provides a ROS2 Humble workspace specifically for the Unitree Go2W wheeled robot, featuring complete 16-motor control drivers, a detailed URDF model, and kinematic simulation tools. It enables developers to interface with the Go2W hardware and simulate its motion within ROS2 environments. The repository is tailored for robotics researchers and engineers working on Unitree's wheeled quadruped platform.

• unitree-sim2real — This project provides a sim-to-sim and sim-to-real pipeline specifically designed for deploying locomotion policies on the Unitree Go2 robot. It leverages Isaac Gym for high-throughput simulation training and includes tools for policy transfer to real hardware, focusing on reinforcement learning-based controllers. The repository targets researchers and engineers working on agile legged locomotion with direct deployment goals on Unitree platforms.

• JeffrinSam_G1 — This project focuses on synthetic joint data generation for Unitree H1 and G1 robots using imitation learning within Isaac Lab. It leverages Isaac Sim for simulation and visualization, enhances realism with NVIDIA Cosmos, and integrates teleoperation data to fine-tune the Gr00t N1.5 model for pick-and-place tasks, with inference validated in Isaac Sim.

• unitree-g1-docker-sdk — This project provides a ready-to-use Docker environment tailored for developing with the Unitree G1 humanoid robot, bundling the official Python SDK, Cyclone DDS middleware, and example code. It enables cross-platform development on Windows and macOS by containerizing dependencies, simplifying setup for G1-specific communication and control. The repository targets developers seeking a streamlined, reproducible environment for Unitree G1 robotics applications.

• butler-connect — Butler-Connect is a professional control system tailored for the Unitree Go2 quadruped, offering a web-based interface for real-time command, safety monitoring, and operational oversight. Built in Python, it enables remote interaction with the robot while enforcing safety protocols, making it suitable for developers and operators deploying Go2 in field applications.

• Unitree_Go2_gallop_Checkpoint — This project provides an Isaac Lab checkpoint specifically trained for galloping locomotion on the Unitree Go2 quadruped robot. It leverages NVIDIA's Isaac Lab simulation framework to enable high-speed dynamic movement policies that can be deployed or fine-tuned for real-world Go2 hardware. The checkpoint serves as a starting point for researchers and developers working on agile legged locomotion using reinforcement learning in simulation.

• isaac-g1-ulc-vlm — This project implements a neuro-symbolic AI framework for the Unitree G1 humanoid robot using Isaac Lab, combining Vision-Language Models (VLMs) with reinforcement learning for task reasoning and control. It leverages ULC (Universal Locomotion Controller) and PPO-PyTorch to enable VLM-guided policy learning in simulation, targeting Sim2Real transfer. Designed for researchers exploring cognitive robotics and embodied AI on the Unitree G1 platform.

• g1_spinkick_example — This project demonstrates how to train and deploy a double spin kick motion on the Unitree G1 humanoid robot using MuJoCo-based simulation with mjlab and mujoco-warp. It provides training data, pretrained models, and Sim2Real deployment instructions tailored specifically for the G1 platform. The repository is aimed at researchers and developers working on dynamic locomotion and whole-body control for Unitree's humanoid robots.

• humanoid-robot-vla-study — This research project evaluates NVIDIA's Isaac GR00T N1.5 vision-language-action (VLA) model on the Unitree G1 humanoid robot, empirically analyzing discrepancies between vendor performance claims and real-world deployment results. It provides hardware-specific insights into VLA model limitations on the G1 platform, including latency, robustness, and task execution fidelity. The study is valuable for researchers and engineers working on deploying foundation models on Unitree humanoid hardware.

• sim2real_go2_rl — This project provides a complete Sim2Real pipeline for deploying Reinforcement Learning policies on the Unitree Go2 Edu quadruped robot. It integrates ROS2 Humble, Linux 22.04, and Unitree SDK2 Python to bridge simulation-trained policies to real hardware execution. The repository targets researchers and developers working on legged locomotion with direct support for the Go2 platform.

• unitree-g1-rl — This project focuses on applying reinforcement learning (RL) to the Unitree G1 humanoid robot, aiming to develop agile and adaptive motor control policies. It likely includes simulation environments and training frameworks tailored for the G1's kinematics and dynamics, potentially leveraging platforms like Isaac Gym or MuJoCo for scalable RL training. The repository is intended for researchers and developers working on embodied AI and humanoid locomotion using the Unitree G1 platform.

• unitree-g1-rl-unity — This project provides a reinforcement learning pipeline specifically designed for the Unitree G1 humanoid robot using Unity and ML-Agents. It features URDF-based simulation, implementations of PPO and SAC algorithms, training configurations, model checkpoints, and experimental results focused on humanoid locomotion tasks. The repository targets researchers and developers working on RL-based control for the Unitree G1 platform.

• humanoid_shadow_g1 — This project integrates the Unitree G1 humanoid robot with Shadow Dexterous Hands to create a full-body model for advanced dexterous manipulation tasks. Built as an extension of the MuJoCo Menagerie, it enables high-fidelity simulation of complex hand-arm coordination in the MuJoCo physics engine. The repository targets researchers and developers working on humanoid manipulation and Sim2Real transfer using the Unitree G1 platform.

• FAST-LIVO2 — FAST-LIVO2 is a tightly-coupled LiDAR-inertial-visual odometry system designed for real-time state estimation on legged robots. This fork specifically targets the Unitree G1 platform, integrating its sensor suite including stereo cameras and IMU for robust navigation in dynamic environments. The implementation leverages C++ for efficiency and is tailored to the G1's hardware interface, making it suitable for researchers and developers working on autonomous locomotion with Unitree's humanoid robot.

• G1_xr_deployment — This project implements a VR-based teleoperation and data collection system specifically designed for the Unitree G1 humanoid robot. It enables immersive control through virtual reality hardware and facilitates dataset generation for downstream learning tasks, leveraging the G1's API for real-time command execution. The system is aimed at researchers and developers working on human-in-the-loop robot learning and telepresence applications.

• Humanoid-Locomotion-via-primitive-composition — This project presents a reinforcement learning approach for humanoid locomotion on the Unitree H1 robot by composing primitive policies to achieve robust bipedal walking. It leverages policy composition techniques in simulation, likely using Isaac Gym or MuJoCo, and targets sim-to-real transfer for dynamic walking behaviors. The work is aimed at researchers and developers focused on agile locomotion control for Unitree's H1 humanoid platform.

• unitree-go2 — This repository provides hardware-ready deep reinforcement learning (DRL) implementations specifically designed for the Unitree Go2 quadruped robot. It includes trained policies and deployment scripts that enable real-world locomotion control on the Go2 platform, leveraging Python-based interfaces to interact with the robot's low-level SDK. The project targets researchers and developers working on DRL for legged locomotion who seek to deploy algorithms directly on Unitree hardware.

• BOLT — B.O.L.T is a real-time vision-guided object tracking system specifically designed for the Unitree Go2 quadruped. It integrates a custom YOLOv8 model with monocular depth estimation and a state-based controller to enable autonomous detection, tracking, and following of objects with sub-50 ms latency on edge hardware. The project targets robotics researchers and developers working on perception-driven locomotion for Unitree Go2.

• Isaaclab_deploy — This project focuses on deploying reinforcement learning (RL) policies trained in Isaac Lab onto the Unitree Go2 robot, enabling real-world execution of simulated locomotion or manipulation behaviors. It provides integration tools and deployment scripts specifically tailored for the Go2 platform, bridging the sim-to-real gap using Isaac Lab as the training environment. The repository is aimed at researchers and engineers working on RL-based control for Unitree’s quadrupedal robots.

• Go2Testing — This project provides an IsaacLab extension specifically designed for developing and testing control algorithms on the Unitree Go2 quadruped robot. It leverages NVIDIA's IsaacLab simulation framework to enable rapid prototyping, sensor integration, and policy deployment with direct support for Go2's hardware interface and kinematic structure. The tool is aimed at researchers and developers working on legged locomotion and Sim2Real transfer for Unitree robots.

• quadruped-locomotion-deploy — This project provides a deployment system for executing trained reinforcement learning policies on the Unitree Go2 quadruped robot. It bridges simulation-trained models to real-world hardware, enabling agile locomotion control through low-level motor commands. The system is built in Python and targets researchers and engineers working on Sim2Real transfer for legged robotics.

• Go2-Locomotion-Hackathon — This project focuses on training locomotion policies for the Unitree Go2 quadruped robot using reinforcement learning in the Genesis simulator. It implements diverse motion skills including walking, running, spinning, and dancing through RL algorithms. The repository provides a complete pipeline from simulation training to potential real-world deployment, targeting researchers and developers working on agile legged locomotion.

• go2-ubuntu22-jetpack6-upgrade — This project provides documentation and shell scripts to upgrade the Unitree Go2 EDU robot's operating system from Ubuntu 20.04 to 22.04 while integrating NVIDIA JetPack 6.2.1 (L4T R36.4.4). It specifically addresses compatibility and setup challenges for the Go2 platform, ensuring proper driver and runtime support on the newer OS and Jetson environment. The repository is targeted at developers and researchers maintaining or modernizing Go2 hardware with updated software stacks.

• butler-connect — Butler-Connect is a professional control system designed specifically for Unitree Go2 quadruped robots, offering a web-based interface for real-time control, safety monitoring, and operational oversight. It enables remote interaction with the Go2 through a browser, integrating hardware-level command interfaces and safety protocols tailored to Unitree's platform. The project targets developers and operators seeking a user-friendly yet robust solution for deploying and managing Go2 robots in real-world applications.

• LISA-V2 — LISA-V2 is an autonomous navigation framework implemented on the Unitree Go2W quadruped robot, enabling on-device perception and planning with optional offloading of large language model (LLM) inference to a co-located edge server. The system integrates real-time sensor processing, navigation stacks, and modular LLM-based decision-making tailored for the Go2 platform. It targets researchers and developers building embodied AI applications on Unitree robots.

• GR00T-training-pipeline-for-Unitree-G1 — This project focuses on fine-tuning NVIDIA's GR00T N1.5 foundation model for a custom manipulation task specifically on the Unitree G1 humanoid robot. It includes setting up a task scene, collecting demonstrations, fine-tuning the model, and deploying checkpoints in both simulation and real-world environments. The work targets researchers and developers aiming to adapt large-scale robot learning models to Unitree's G1 platform.

• motion_tracking_controller — This ROS 2 Humble package enables deployment of reinforcement learning policies to the Unitree G1 humanoid robot, supporting both MuJoCo simulation via LeggedGym/unitree_rl_gym and real-world execution. It provides a unified sim-to-real control interface, loads policies from WandB or local ONNX files, and publishes real-time actions to the robot. The project is tailored for researchers and developers working on Unitree-G1-specific locomotion and motion tracking using RL-based whole-body control.

• g1 — This project enables simulation, training, and deployment of locomotion policies specifically for the Unitree G1 humanoid robot using the Genesis framework on Mac M-series hardware. It leverages Python-based reinforcement learning pipelines tailored to the G1's kinematics and dynamics, allowing developers to iterate on control policies locally before real-world deployment. The repository targets researchers and engineers working on humanoid locomotion with Unitree's G1 platform.

• g1-motion-control — This project provides an engineering-oriented motion control and imitation learning pipeline specifically designed for the Unitree G1 humanoid robot. It implements command-conditioned crawling and full-body motion tracking, leveraging Python-based control algorithms to enable dynamic locomotion and pose replication. The repository targets robotics developers and researchers working on whole-body control and Sim2Real transfer for the Unitree G1 platform.

• G1_Userguide_for_lab_of_FSII — This project provides a comprehensive developer-oriented manual for the Unitree-G1 robot, tailored for FSII lab use. It covers mechanical structure, operational guidelines, application development, low-level motion control, and high-level motion programming, with content adapted from official Unitree documentation. The guide serves as a practical resource for researchers and engineers working directly with the Unitree-G1 platform.

• unitree-h1-ros-sdk — This project provides a ROS2 SDK specifically designed for the Unitree H1 humanoid robot, enabling developers to interface with the robot's hardware through ROS2 nodes. It includes C++ implementations for low-level control, sensor data access, and command publishing, targeting researchers and engineers working on humanoid robotics applications with the H1 platform.

• LearningHumanoidWalking_H1 — This project focuses on training the Unitree H1 humanoid robot for locomotion using Reinforcement Learning. It implements RL-based policies to enable stable walking behaviors, likely leveraging simulation environments such as Isaac Gym or MuJoCo for training before potential real-world deployment. The repository is tailored specifically for the Unitree-H1 platform, making it relevant for researchers and developers working on humanoid locomotion.

• unitree_h12_sim2sim — This project provides a comprehensive pipeline for training locomotion policies in IsaacLab and performing sim-to-sim (Sim2Sim) transfer to MuJoCo, specifically targeting the Unitree H1-2 humanoid robot. It includes environment setup, policy training scripts, and domain adaptation tools to bridge simulation frameworks. The repository is aimed at researchers and developers working on humanoid robot control and Sim2Real transfer for Unitree platforms.

• unitree_g1_vibes — This project provides Python-based tools and examples for interacting with the Unitree G1 humanoid robot, focusing on motion control and real-time command interfaces. It includes utilities for sending joint commands and managing robot states, likely leveraging Unitree's official SDK. The repository is aimed at developers and researchers working on humanoid robotics applications with the G1 platform.

• go2_dashboard — This project provides a simple web-based dashboard for monitoring and interacting with the Unitree Go2 robot. Built in Python, it offers real-time visualization of robot states such as joint angles, IMU data, and battery status through a browser interface. It is designed for developers and researchers who need quick access to Go2 telemetry without complex setup.

• AIM-Robotics — AIM-Robotics is a Python-based project focused on the Unitree G1 humanoid robot, providing tools and examples for control, perception, and autonomy. It includes implementations for whole-body motion planning and sensor integration tailored specifically for the G1 platform. The repository targets researchers and developers working with Unitree's G1 robot for advanced robotics applications.

• Aliengo_2D_Nav-sim — This project implements 3D SLAM and 2D navigation capabilities specifically for the Unitree Aliengo quadruped robot using C++. It integrates ROS-based navigation stacks with robot-specific configurations to enable autonomous indoor navigation in simulated environments. The repository targets researchers and developers working on legged robot autonomy who need a baseline for perception and navigation on the Aliengo platform.

• unitree_g1_ros2_demo — This project provides a ROS 2 demonstration for the Unitree G1 humanoid robot, showcasing basic control and interaction capabilities using Python. It includes example nodes for robot state visualization and command publishing, leveraging the official Unitree SDK through ROS 2 interfaces. The repository serves as a starting point for developers exploring ROS 2 integration with the Unitree G1 platform.

• unitree_h1_humanoid-gym — This project implements a reinforcement learning environment for the Unitree H1 humanoid robot using the Humanoid-Gym framework. It provides a customized RL setup with Isaac Gym integration, enabling policy training and simulation specifically tailored for the H1's kinematics and dynamics. The repository targets researchers and developers working on humanoid locomotion and control for Unitree H1.

• Unitree-Go2-GestureDetection — This project enables gesture recognition for the Unitree Go2 robot dog using MediaPipe and the official Python SDK, allowing the robot to respond to hand gestures in real time. It integrates computer vision with robot control, providing a natural human-robot interaction interface. The implementation is tailored specifically for Go2’s hardware and motion capabilities, making it useful for developers exploring intuitive teleoperation methods.

• The-way-of-redemption-lies-within — This project implements Model Predictive Control (MPC) specifically for the Unitree G1 humanoid robot. It provides low-level control algorithms tailored to G1's hardware and dynamics, likely interfacing with Unitree's SDK or real-time control interfaces. The repository targets researchers and developers working on advanced motion control for the G1 platform.

• Unitree-G1-Control-Replay — This project provides a control and replay platform specifically designed for the Unitree G1 humanoid robot, enabling users to record, visualize, and replay motion trajectories. It features Python-based interfaces for real-time control and data logging, leveraging the robot's native SDK for low-level command execution. The platform is tailored for researchers and developers working on humanoid locomotion and motion planning with the Unitree G1.

• ark_unitree_g1 — This project integrates the Unitree G1 humanoid robot with the Ark robotics framework, enabling high-level control and task execution through Ark's modular architecture. It provides Python-based interfaces for commanding G1's actuators and sensors, leveraging Ark's real-time capabilities for whole-body control and perception tasks. The repository targets researchers and developers working with Unitree G1 who seek to build complex robotic behaviors using Ark.

• unitree-g1-brainco-hand — This project provides a tutorial for integrating BrainCo Revo2 prosthetic hands with the Unitree G1 humanoid robot. It includes Python-based control examples and hardware interfacing guidance specifically tailored for the G1 platform, enabling dexterous hand manipulation capabilities. The repository serves robotics researchers and developers working on enhancing Unitree G1's upper-body functionality with advanced prosthetic hands.

• isaac-b2-ros2 — This project provides a ROS 2 Humble integration for the Unitree B2 robot within the Isaac Sim environment, enabling simulation and control workflows. It includes Python-based interfaces to connect Isaac Sim with ROS 2 topics, facilitating sensor data publishing and command execution specific to the Unitree B2 platform. The repository targets developers working on legged locomotion and robotics research using NVIDIA's Isaac Sim and Unitree’s B2 hardware.

• go2_description — This project provides a ROS 2 Humble-compatible robot description package for the Unitree Go2, built on top of unitree_ros and tailored for integration with the CHAMP controller. It includes URDF models and configuration files enabling simulation and control workflows in ROS 2. The repository is actively maintained and specifically targets developers working with Unitree Go2 in ROS-based robotics applications.

• Unitree-Go2-Mapping-and-Navigation-Using-Intel-RealSense-D435i-and-RTAB-Map — This project implements mapping and navigation capabilities for the Unitree Go2 robot using an Intel RealSense D435i depth camera and RTAB-Map for SLAM. It integrates ROS-based perception and localization pipelines tailored to the Go2's hardware, enabling autonomous navigation in indoor environments. The repository targets robotics developers working on legged robot autonomy with off-the-shelf depth sensors.

• Quadruped-Robot-State-Estimation-Go2-Kinematics-Velocity-Tracking — This project implements real-time state estimation for the Unitree Go2 quadruped using forward kinematics and sensor fusion to estimate position, orientation, and velocity from onboard sensors. It provides a practical solution for autonomous navigation by leveraging Go2-specific kinematics and IMU data, targeting robotics developers working on legged locomotion and perception.

• Unitree-G1-Mujoco-Playground — This repository provides a MuJoCo-based simulation environment for the Unitree G1 humanoid robot, enabling users to develop and test control policies in a physics-accurate simulator. It includes model definitions, basic controllers, and example scripts tailored specifically for the G1 platform. The project targets researchers and developers working on humanoid locomotion and whole-body control using MuJoCo.

• Go2Bot-OpenAI-Integration — This project integrates the Unitree Go2 robot with OpenAI to enable voice-command processing and AI-driven task execution, enhancing human-robot interaction. Built with Python, Flask, and JavaScript, it demonstrates a practical application of large language models for controlling a Unitree Go2 in real-world scenarios. The repository is aimed at researchers and developers exploring intelligent interfaces for quadruped robots.

• go2-vision-head — This project implements a vision-based head system specifically for the Unitree Go2 series (AIR/PRO/EDU), enabling camera-driven perception and head movement control. It integrates with the Go2's hardware interface to provide real-time visual feedback and active head orientation, useful for teleoperation or autonomous navigation tasks. The repository targets developers working on perception-enhanced applications for Unitree's quadruped platform.

• unitree_deploy — This project deploys locomotion policies from the Ctrl Lab onto the Unitree Go2 robot, enabling real-world execution of learned control strategies. It provides Python-based deployment scripts that interface directly with the Go2's hardware, facilitating Sim2Real transfer for agile locomotion tasks. The repository is aimed at researchers and developers working on legged robot control using Unitree's Go2 platform.

• Unitree_G1_Stabilization — This project provides a Python-based stabilization implementation for the Unitree G1 humanoid robot using the Pinocchio library for rigid-body dynamics. It focuses on maintaining balance and posture control through inverse kinematics and center-of-mass tracking, leveraging the official G1 URDF model. The repository targets researchers and developers working on humanoid locomotion and whole-body control for the Unitree G1 platform.

• g1_description — This ROS2 package provides the robot description for the Unitree G1 humanoid robot, including URDF files and associated launch configurations. It enables integration with ROS2-based simulation and control frameworks by defining the robot's kinematic and visual properties. The project is aimed at developers working on G1-specific applications in ROS2 environments.

• go2_ros2_control_sim — This project provides a ROS 2 control implementation for the Unitree Go2 quadruped robot, supporting both classic control and reinforcement learning (RL)-based control strategies. It integrates with ROS 2 control frameworks and is designed for simulation environments, enabling developers to test and deploy locomotion controllers. The repository targets robotics researchers and engineers working on legged locomotion with the Unitree Go2 platform.

• go2webrtc-rs — This project provides a Rust implementation for WebRTC streaming from the Unitree Go2 robot, enabling real-time video and data transmission. It directly interfaces with the Go2's onboard systems to establish peer-to-peer communication using WebRTC protocols. The tool is aimed at developers building remote monitoring or teleoperation applications for the Unitree Go2 platform.

• Unitree-Go2-EDU-RL-Training-and-Deployment-Tutorial — This repository provides an educational tutorial for training and deploying reinforcement learning (RL) policies on the Unitree Go2 robot. It includes practical examples and workflows for RL algorithm implementation, simulation-to-real (Sim2Real) transfer, and deployment using frameworks compatible with the Go2 platform. The project is aimed at students and researchers new to legged locomotion and RL on real-world quadruped robots.

• unitreeG1_ik — This project provides inverse kinematics (IK) solutions specifically for the Unitree G1 humanoid robot, enabling precise end-effector pose control through analytical or numerical methods. Implemented in Python, it offers a lightweight and accessible tool for developers working on motion planning or teleoperation tasks with the G1 platform. The repository is tailored to Unitree G1's kinematic chain and joint limits, making it directly relevant for researchers and engineers building on this hardware.

• isaacsim_g1_locomotion — This project provides a standalone Python script for simulating Unitree G1 locomotion within NVIDIA Isaac Sim. It enables developers to test and develop locomotion controllers for the G1 humanoid robot in a high-fidelity simulation environment, leveraging Isaac Sim's physics and rendering capabilities. The repository is aimed at robotics researchers and engineers working on bipedal locomotion for the Unitree G1 platform.

• unitree_g1_primitives — This project provides an easy-to-setup motion primitives library specifically designed for the Unitree G1 humanoid robot. It enables developers to quickly implement and test basic locomotion and manipulation behaviors using Python, leveraging the robot's native control interfaces. The library is aimed at robotics researchers and engineers working on high-level behavior development for the Unitree G1 platform.

• unittree-go2-usd — This project provides a USD (Universal Scene Description) scene file for the Unitree Go2 quadruped robot, enabling its use in NVIDIA Isaac Sim for high-fidelity simulation and development. It includes properly configured articulation, collision geometries, and visual assets tailored for the Go2 platform, facilitating tasks like motion planning, control policy testing, and sensor simulation. The repository is aimed at robotics researchers and developers leveraging Isaac Sim for Unitree Go2-specific applications.

• OpenGo2Air — OpenGo2Air is an open-source hardware and software project built upon the Unitree Go2 quadruped platform, aiming to provide a fully transparent and modifiable alternative for research and development. It includes custom PCB designs, motor drivers, and embedded firmware tailored specifically for the Go2's form factor and dynamics. The project targets robotics researchers and hobbyists seeking low-level access and hardware-software co-design capabilities on a Unitree-based platform.

• exploration_go2 — This project implements a frontier-based SLAM and autonomous exploration system specifically designed for the Unitree Go2 quadruped robot. Built on ROS2 and Nav2, it enables the Go2 to autonomously navigate and map unknown environments using standard robotic perception and planning stacks. The repository provides integration with the Go2's hardware interface and is aimed at researchers and developers working on legged robot autonomy.

• QRC_ARCLAB_Code — This project modifies the Unitree Go2 SDK to support the QRC2024 competition, providing customized low-level control and communication interfaces for the Go2 robot. It includes C++ implementations tailored for real-time locomotion and sensor integration specific to competition tasks. The repository is aimed at robotics developers and competition participants working with Unitree Go2 hardware.

• foot_reach — This project implements a simple foot target reaching task for the Unitree Go2 quadruped robot using IsaacLab, a high-performance reinforcement learning and robotics simulation framework. It provides a minimal yet functional example of whole-body control and task-space trajectory tracking specifically tailored for the Go2's kinematics and dynamics. The repository serves as a starting point for researchers and developers exploring legged locomotion control or Sim2Real transfer on Unitree platforms.

• huro — huro provides a ROS2 interface specifically designed for Unitree robots, enabling seamless integration with ROS2 ecosystems. It includes C++ implementations for robot control, state feedback, and sensor data handling tailored to Unitree's hardware protocols. This project is aimed at developers and researchers building applications on Unitree platforms using ROS2.

• go2_rl_gym — This repository provides a reinforcement learning implementation specifically designed for the Unitree Go2 quadruped robot. It includes custom Gym environments and training frameworks tailored to Go2's dynamics and control interface, enabling researchers to develop and test RL policies for locomotion and other tasks. The project targets robotics developers and researchers working on agile legged locomotion using the Unitree Go2 platform.

• g1_description — This project provides an optimized URDF description for the Unitree G1 humanoid robot, enhanced with a joint state publisher to enable real-time visualization in RViz. It focuses on accurate kinematic representation and seamless integration with ROS tools for simulation and debugging. The repository is useful for developers and researchers working with the Unitree G1 in ROS-based environments.

• Unitree-G1-Research — This repository focuses on research and engineering efforts specifically for the Unitree G1 humanoid robot, providing Python-based tools and experimental implementations. It appears to support development workflows tailored to the G1 platform, potentially including control, perception, or simulation components. The project targets researchers and engineers working directly with the Unitree G1.

• g1_hardware — This project provides a C++ hardware interface specifically designed for controlling the arms of the Unitree G1 humanoid robot. It enables low-level communication with the G1's arm actuators, facilitating real-time control and integration into custom robotic applications. The repository is targeted at developers and researchers working directly with the Unitree G1 platform who need reliable hardware-level access.

• Voice-Interaction-Control-on-Unitree-G1-Robot-to-realize-FALCON- — This project integrates microphone input, a large language model (LLM), and the FALCON framework to enable voice-interaction control on the Unitree G1 robot. It implements real-time audio capture, natural language command processing via an LLM, and translates high-level instructions into robot actions using FALCON's motion generation pipeline. The system is built in C++ and targets