GalaxyGeneralRobotics/Humanoid-GPT

What it solves

Humanoid-GPT addresses the limitations of traditional humanoid motion tracking, which often rely on shallow models and small datasets. This typically leads to a trade-off between how agile a robot can be and how well it generalizes to new movements. This project enables zero-shot motion tracking, allowing a humanoid robot to perform arbitrary, unseen motions without needing specific fine-tuning for each new task.

How it works

The system uses a GPT-style Causal Transformer architecture equipped with Rotary Position Embeddings (RoPE) to handle variable-length motion sequences. It was pre-trained on a massive scale—2 billion motion frames—by unifying various motion capture (mocap) datasets and in-house recordings. The model is optimized for the Unitree G1 humanoid robot (29 degrees of freedom) and utilizes GPU-accelerated simulation via MuJoCo-MJX for efficient evaluation and training.

Who it’s for

This project is designed for robotics researchers and engineers working on whole-body control for humanoid robots, specifically those using the Unitree G1 platform or similar high-DOF humanoid systems.

Highlights

• Billion-Scale Training: Pre-trained on a corpus of 2 billion motion frames.
• Zero-Shot Generalization: Ability to track unseen motions without additional training.
• GPT-Style Architecture: Uses a Causal Transformer with RoPE for flexible sequence handling.
• Hardware Ready: Includes deployment modules for the Unitree G1, including on-board Jetson Orin support.