This project addresses zero-shot motion tracking for humanoid robots by introducing Humanoid-GPT, a novel GPT-style motion Transformer. It targets researchers and engineers in robotics and AI, offering unprecedented generalization capabilities to track arbitrary unseen motions without task-specific fine-tuning, significantly advancing whole-body control.

How It Works

Humanoid-GPT employs a causal Transformer architecture with Rotary Position Embeddings (RoPE), enabling it to process variable-length motion sequences. It is pre-trained on an unprecedented 2 billion motion frames, consolidating diverse motion capture datasets with large-scale in-house recordings. This massive data scale, combined with the Transformer's capacity, allows for superior generalization compared to prior shallow MLP trackers, overcoming the agility-generalization trade-off.

Quick Start & Requirements

• Primary install command: After cloning the repository and creating a Python 3.12 conda environment (conda create -n h-gpt python=3.12 -y), install with pip install -e ".[cuda]" (or .[cpu] on MacOS, or . for real robot deploy-only).
• Non-default prerequisites: NVIDIA GPU with CUDA 12.x is required for GPU acceleration. MacOS users can use mjpython for MuJoCo viewer functionality.
• Hardware: The platform is optimized for the Unitree G1 humanoid robot (29 DOF whole-body).
• Links: The project is associated with the CVPR 2026 paper "Humanoid-GPT: Scaling Data and Structure for Zero-Shot Motion Tracking".

Highlighted Details

• Billion-Scale Pre-Training: First project to scale humanoid motion learning to 2 billion frames, unifying major mocap datasets.
• Zero-Shot Generalization: Achieves unprecedented generalization to unseen motions and tasks without requiring fine-tuning.
• GPT-Style Architecture: Utilizes a causal Transformer with RoPE, supporting variable-length motion sequences.
• Platform Optimization: Specifically optimized for the Unitree G1 humanoid robot.

Maintenance & Community

The provided README does not contain information regarding notable contributors, sponsorships, community channels (e.g., Discord, Slack), or a public roadmap.

Licensing & Compatibility

• License type: Apache 2.0.
• Compatibility notes: Built upon MuJoCo, Brax, and the Unitree G1 platform. The Apache 2.0 license is generally permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Core inference and deployment code are marked as "TODO" for future implementation. Pre-trained model checkpoints are also listed as "TODO", although a specific ONNX path is provided. The project is primarily optimized for the Unitree G1 humanoid robot, suggesting potential limitations for other hardware platforms without adaptation.