galaxygeneralrobotics/humanoid-gpt

Humanoid-GPT is a research project from CVPR 2026 focused on teaching humanoid robots to mimic human body movements. The core idea is motion tracking: given a recording of a person moving, the system controls a humanoid robot to reproduce that motion as accurately as possible. The approach borrows a pattern from large language models. Just as GPT-style language models are pre-trained on enormous amounts of text before being applied to specific tasks, this system is pre-trained on a very large collection of motion capture data, which is recordings of human movement captured by sensors or cameras. The training dataset contains around two billion frames of motion data drawn from major publicly available motion capture collections combined with additional recordings. The model uses an architecture called a causal Transformer, which is the same type of structure underlying many modern language models, adapted here to process sequences of body poses over time. The claimed benefit of training at this scale is zero-shot generalization, meaning the system can track new types of movement it was never explicitly trained on, without needing to be fine-tuned for each new motion type. This addresses a known problem in the field where smaller models trained on limited data tend to either specialize narrowly or generalize poorly. The hardware target is the Unitree G1, a commercially available humanoid robot with 29 independently controllable joints covering the whole body including arms, legs, and torso. The repository includes a pre-trained model file and example motion data so you can run inference immediately after installing dependencies. There is also an interactive web demo built with Gradio, a tool for creating simple browser interfaces for machine learning models. Deployment instructions are included for running on the actual physical robot, including configuration for the onboard computer that rides on the robot and a variant that uses a connected hand device. Training code and training data are listed as planned additions but are not yet in the repository. The project is licensed under Apache 2.0.