BlackCatRoboticsAI/g1-amp-motion

G1-AMP Motion Imitation Policy

Unitree G1 Humanoid Motion Imitation (AMP) — 494 human motion capture sequences from the AMASS dataset, trained using Adversarial Motion Priors (AMP) in Isaac Lab.

Model Details

Field	Value
License	MIT (commercial use ✅)
Robot	Unitree G1 (37 DOFs, 23 active)
Input	216-dim observations (joint pos/vel, root state, future reference targets)
Output	23-dim joint position targets (scale by 0.5)
Framework	Isaac Lab / PyTorch
Format	TorchScript (policy_jit.pt = 2.9MB)

Performance

Metric	Value
Total Reward (mean)	160.29
Episode Length (mean)	396.9 / 400 steps
Tracking Reward (mean)	0.317

Quick Start

import torch
import numpy as np

# Load policy
model = torch.jit.load("policy_jit.pt")
model.eval()

# Run inference
obs = torch.randn(1, 216)  # Replace with actual observations
with torch.no_grad():
    actions = model(obs)  # (1, 23) joint position targets
    actions = actions * 0.5  # Scale before sending to robot

Files

• policy_jit.pt — JIT-traced policy for deployment (2.9MB)
• best_agent.pt — Full checkpoint for resuming training (25MB)

Training Details

• Motion Dataset: 494 AMASS motions (~113 min, 196,642 frames)
• Algorithm: AMP (Adversarial Motion Priors) via skrl
• Training Hardware: NVIDIA RTX 4080 SUPER (16GB VRAM)
• Training Duration: 5.5 days (12.4M timesteps)
• Domain Randomization: Mass, friction, PD gains, action delay

Citation

@misc{pathonai2026g1imitate,
  title={G1 Humanoid Motion Imitation with AMP in Isaac Lab},
  author={PathOn-AI},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/PathOn-AI/g1-imitate-isaaclab-amp}
}

License

MIT License — free for commercial and non-commercial use.