Genesis Humanoid

Genesis Humanoid is an all-in-one humanoid research platform on top of the Genesis simulator. It supports real-time human-to-humanoid retargeting, includes different motion dataset with unified format, provides an end-to-end learning pipeline, and offers a modular low-level control testbed for experimentation.

Outline

• Features
• Code Structure
• Code Style
• Installation
• Usage
  • Process humanoid motions
  • Retarget human motion
  • Run RL training
  • Run BC distillation
  • RL finetune / resume training
  • Evaluate trained policy
  • Deploy to a real robot
  • Teleoperate the robot
• Unified Motion Dataset
• Citation
• Acknowledgement

Supported Hardware

• OptiTrack Mocap src/env/gs_env/real/optitrack
• SteamVR src/env/gs_env/real/steamvr
• ChangingTek Gripper (CTAG2F90-D) src/env/gs_env/real/changingtek
• RealSense Camera (Async streaming by shared memory) src/env/gs_env/real/realsense
• WitMotion IMU (WT901C) src/env/gs_env/real/witimu

Features

Genesis Humanoid shares the modular design with GenesisPlayground, featuring with:

• macOS compatibility provides an interactive viewer on your MacBook.

• Sim-Real duality enables seamless sim2real deployment with the same code.

• Extraordinary speed that reaches 200k RL steps per second (8192 environments, decimation 4 and 29 articulated joints, tested on a NVIDIA L40s with run_ppo_walking.py), which is nearly 0.8M FPS.

We list several featured projects built on Genesis Humanoid below:

ExtremControl A whole-body teleoperation system for Unitree G1 with 50ms end-to-end latency.

Code Structure

Adopting from GenesisPlayground, Genesis Humanoid is built with:

• gs-schemas - Shared data structures and interfaces
• gs-agent - Robot learning algorithms (PPO and DAgger)
• gs-env - Wrapped environments including both simulation and real world
• examples - Ready-to-run examples in simulation
• deploy - Ready-to-run examples in the real world

To setup a new simulation environment, write new robot and environment configurations in src/env/gs_env/sim/robots/config/registry.py & src/env/gs_env/sim/envs/config/registry.py.

Code Style

We use pre-commit to enforce code formatting and linting.

pre-commit install
pre-commit run --all-files

Installation

1. Install Dependencies

# Install uv (fast Python package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and setup the repository
git clone --recurse-submodules git@github.com:UMass-Embodied-AGI/Genesis-Humanoid.git
cd Genesis-Humanoid

2. Install the gs-env package

# uv
uv sync --package gs-env

# pip
uv pip compile pyproject.toml -o requirements.txt --python /PATH/TO/PYTHON
python -m pip install -r requirements.txt

3. Activate the environment

source .venv/bin/activate

4. Setup real-world environment

Please refer to READMEs in each folder under src/env/gs_env/real.

Usage

A wide range of example usages of Genesis Humanoid can be found in the /examples and /deploy directories.

Process humanoid motions

# LAFAN1 (Optional but recommended)
python examples/convert_lafan.py

# HuB (Optional)
python examples/convert_hub.py

Retarget human motion

# Recorded MoCap motion
python examples/convert_optitrack.py

# AMASS (Optional)
# Download SMPLX body model to assets/body_models
# Download AMASS dataset to assets/AMASS
python examples/convert_amass.py

For motion retargeting, Genesis Humanoid integrates Cartesian-spacing mapping from ExtremControl and cleans up GMR for joint-space retageting. Check Teleoperate for real-time retargeting from OptiTrack or SteamVR.

Run RL training

# Teleop teacher policy
python examples/run_ppo_motion.py \
    --exp_name TEACHER_EXP_NAME \
    --env_name g1_motion_teacher \
    --env.motion_file assets/motion/motion.yaml

Run BC distillation

# Distill a deployable policy
python examples/run_bc_motion.py \
    --exp_name BC_EXP_NAME \
    --env_name g1_motion \
    --teacher_exp_name TEACHER_EXP_NAME \
    --env.motion_file assets/motion/motion.yaml

RL finetune / resume training

# Finetune the distilled policy
python examples/run_ppo_motion.py \
    --exp_name BC_EXP_NAME \
    --env_name g1_motion \
    --resume True \
    --use_stored_config False \
    --runner.freeze_actor_iterations 200 \
    --algo.lr 3e-5 \
    --env.motion_file assets/motion/teacher.yaml

Evaluate trained policy

# Evaluation will store a deployable policy to deploy/logs/EXP_NAME
python examples/run_ppo_motion.py \
    --exp_name EXP_NAME \
    --num_ckpt NUM_CKPT (optional) \
    --eval True \
    --show_viewer True \
    --env.motion_file assets/motion/evaluate.pkl

Deploy to a real robot

Install unitree-sdk2-python for deployment. Install redis for teleoperation.

# Make sure deploy/logs/EXP_NAME exists
# Sanity check in simulation
uv pip install redis
python deploy/g1_teleop.py --exp_name EXP_NAME
python deploy/g1_motion.py \
    --exp_name EXP_NAME \
    --motion_file MOTION_PATH

# Deployment should start from small ACTION_SCALE
python deploy/g1_teleop.py \
    --exp_name EXP_NAME \
    --sim False \
    --action_scale ACTION_SCALE
python deploy/g1_motion.py \
    --exp_name EXP_NAME \
    --motion_file MOTION_PATH \
    --sim False \
    --action_scale ACTION_SCALE

Teleoperate the robot

# Test with existing motion
python deploy/motion_publisher.py --motion_file MOTION_FILE

# Optitrack MoCap
python deploy/optitrack_publisher,py

# SteanVR
python deploy/steamvr_publisher.py

Unified Motion Dataset

The convert_[ ].py files convert the existing motions from different format into:

{
    "fps": 50,
    "link_names": ["LINK_NAMES_IN_ORDER"],
    "dof_names": ["DOF_NAMES_IN_ORDER"],
    "pos": torch.Tensor, # [num_frames, 3] position of the root link
    "quat": torch.Tensor, # [num_frames, 4] quaternion of the root link
    "dof_pos": torch.Tensor, # [num_frames, num_dof] dof positions in order
    "link_pos": torch.Tensor, # [num_frames, num_link, 3] link positions in order
    "link_quat": torch.Tensor, # [num_frames, num_link, 4] link quaternions in order
    "foot_contact": torch.Tensor, # [num_frames, 2] foot contact probability
}

Citation

If you find our code useful, please consider citing our related paper:

@misc{xiong2026extremcontrol,
      title={ExtremControl: Low-Latency Humanoid Teleoperation with Direct Extremity Control}, 
      author={Ziyan Xiong and Lixing Fang and Junyun Huang and Kashu Yamazaki and Hao Zhang and Chuang Gan},
      year={2026},
      eprint={2602.11321},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2602.11321}, 
}

Acknowledgement

The entire codebase is built on GenesisPlayground. We thank GMR for serving as a reference for retargeting and OptiTrack streaming. The human datasets used in this project includes AMASS, HuB and LAFAN1.