README.md

metadata

title: G1 Moves
emoji: 🤖
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
license: cc-by-4.0

G1 Moves

Dataset (BVH, FBX, PKL, NPZ, ONNX policies) · Showcase (you are here) · Code (scripts & docs)

60 motion capture clips for the Unitree G1 humanoid robot (29 DOF, mode 15). Dance, karate, fencing, and more — captured with MOVIN TRACIN, retargeted with Motion-Player-ROS, trained with mjlab, and deployed via RoboJuDo.

Browse the interactive gallery above to view trained RL policies running in real-time in your browser via MuJoCo WASM + ONNX Runtime.

---

Pipeline

1. Capture

Motion is captured with the MOVIN TRACIN markerless mocap system (LiDAR + vision, 60 FPS) and recorded in MOVIN Studio. Output is BVH format with a 51-joint humanoid skeleton.

Alternatively, motions can be extracted from any video (YouTube, phone, AI-generated) using video2robot, which runs PromptHMR pose extraction followed by GMR inverse kinematics retargeting — no mocap hardware needed.

2. Retarget

Each BVH clip is retargeted to the Unitree G1's 29-DOF joint space using per-frame inverse kinematics via MOVIN-SDK-Python and a custom ROS 2 retargeting node. The node provides real-time preview with dual BVH skeleton + robot visualization in RViz, and supports both playback of pre-recorded .pkl files and live retargeting from MOVIN TRACIN data via OSC/UDP.

# Playback mode
ros2 launch motion_player player.launch.py \
  motion_file:=clip.pkl bvh_file:=clip.bvh

# Real-time mocap
ros2 launch motion_player realtime.launch.py \
  port:=9000 human_height:=1.75

3. Train

Retargeted motions are trained as reinforcement learning policies using mjlab, which combines Isaac Lab's manager-based API with GPU-accelerated MuJoCo-Warp simulation. PPO training runs across 8,192 parallel environments on a single NVIDIA RTX PRO 6000.

• Actor: 4-layer MLP (160 → 512 → 256 → 128 → 29, ELU activations)
• Observation (160-dim): reference joint pos/vel, 14-body anchor tracking, IMU, proprioception
• Control: 50 Hz, adaptive early stopping when time-out ratio >= 0.95
• Reward: position, orientation, velocity, and collision avoidance shaping across 14 tracked bodies

# Train a motion imitation policy
uv run train Mjlab-Tracking-Flat-Unitree-G1 \
  --env.commands.motion.motion-file clip.npz \
  --env.scene.num-envs 8192 \
  --agent.max-iterations 30000

Every trained policy is exported as an ONNX model (160-dim input, 29-dim output) with baked-in observation normalization.

4. Deploy

Policies are deployed to the physical G1 via RoboJuDo, a plug-and-play sim2real framework with modular controller/environment/policy architecture. Supports seamless switching between MuJoCo sim2sim validation and real hardware deployment via Unitree SDK over Ethernet (192.168.123.x). Hardware E-stop and tethered first-run safety protocol required.

# Sim2sim validation
python scripts/run_pipeline.py --config=g1

# Deploy to physical G1
python scripts/run_pipeline.py --config=g1_real

License

CC BY 4.0 — Experiential Technologies

Full observation vector spec and integration guide in the dataset README.

---

Safety Disclaimer & Limitation of Liability

WARNING: Deploying learned locomotion policies to physical robots is inherently dangerous and can cause serious injury, death, or property damage.

The policies provided in this dataset are trained entirely in simulation and have not been validated for safe real-world operation. Sim-to-real transfer involves unpredictable failure modes including but not limited to:

• Sudden loss of balance and uncontrolled falls
• Uncontrolled high-speed or high-torque limb movements
• Unexpected collisions with people, objects, or the environment
• Hardware damage to actuators, joints, or structural components
• Divergent or unstable behavior not observed in simulation
• Failure to respond to emergency stop commands under certain conditions

Requirements for Physical Deployment

Do not deploy these policies to a physical robot unless you have:

1. Advanced expertise in sim-to-real transfer, robot safety systems, real-time control, and the specific robot platform (Unitree G1)
2. Appropriate physical safety infrastructure including overhead tethers or gantry systems, rigid safety enclosures, hardware emergency stop mechanisms independent of software, and adequate clearance from personnel and obstacles
3. Completed thorough sim-to-sim validation in a separate MuJoCo environment to verify policy behavior before any hardware deployment
4. Conducted a formal risk assessment for your specific deployment environment, considering all potential failure modes and their consequences
5. Ensured all personnel in the vicinity are trained on emergency procedures, aware of the risks, and positioned at a safe distance
6. Verified insurance coverage adequate for robotics research and development activities
7. Compliance with all applicable local, state, and federal regulations governing the operation of autonomous robotic systems

Limitation of Liability

THE POLICIES, MODELS, TRAINING DATA, SOFTWARE, ONNX MODELS, CONFIGURATION FILES, AND ALL OTHER MATERIALS IN THIS DATASET (COLLECTIVELY, THE "MATERIALS") ARE PROVIDED "AS IS" AND "AS AVAILABLE" WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, SAFETY, ACCURACY, RELIABILITY, NON-INFRINGEMENT, OR SUITABILITY FOR ANY PARTICULAR APPLICATION.

IN NO EVENT SHALL EXPERIENTIAL TECHNOLOGIES, ITS AFFILIATES, OFFICERS, DIRECTORS, EMPLOYEES, AGENTS, CONTRIBUTORS, LICENSORS, OR ANY OTHER PARTY INVOLVED IN CREATING, PRODUCING, OR DELIVERING THE MATERIALS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, CONSEQUENTIAL, OR PUNITIVE DAMAGES OF ANY KIND (INCLUDING BUT NOT LIMITED TO PERSONAL INJURY, WRONGFUL DEATH, PROPERTY DAMAGE, LOSS OF USE, LOSS OF DATA, LOSS OF PROFITS, BUSINESS INTERRUPTION, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES, OR ANY OTHER COMMERCIAL OR NON-COMMERCIAL DAMAGES OR LOSSES) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, NEGLIGENCE, TORT, OR ANY OTHER LEGAL OR EQUITABLE THEORY, ARISING OUT OF OR IN ANY WAY RELATED TO THE USE, INABILITY TO USE, OR PERFORMANCE OF THE MATERIALS, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES AND EVEN IF A REMEDY SET FORTH HEREIN IS FOUND TO HAVE FAILED OF ITS ESSENTIAL PURPOSE.

Assumption of Risk and Indemnification

By downloading, accessing, using, modifying, or deploying any part of the Materials, you expressly acknowledge and agree that:

• You assume all risk and full responsibility for any and all consequences arising from the use of the Materials, including deployment to physical hardware
• You have the requisite technical expertise to evaluate the safety and suitability of the Materials for your intended use
• You will indemnify, defend, and hold harmless Experiential Technologies, its officers, directors, employees, agents, and affiliates from and against any and all claims, damages, losses, liabilities, costs, and expenses (including reasonable attorneys' fees) arising from or related to your use of the Materials
• The Materials are intended for research and educational purposes and have not been certified or validated for any safety-critical application
• You are solely responsible for ensuring compliance with all applicable laws, regulations, and safety standards in your jurisdiction

This disclaimer applies to all uses of the Materials, whether in simulation, on physical hardware, or in any derivative works.