Humanoid-GPT: A Breakthrough in Zero-Shot Motion Tracking with Scalable Data and Structure

Humanoid-GPT is introduced as a GPT-style Transformer designed for whole-body control, trained on a billion-scale motion corpus. This model overcomes limitations of previous MLP trackers by utilizing a 2B-frame retargeted dataset that integrates major motion capture datasets with extensive in-house recordings. The scaling of both data and model capacity allows Humanoid-GPT to achieve exceptional zero-shot generalization to unseen motions and control tasks. Experimental results demonstrate that the model sets a new performance benchmark, effectively tracking complex and dynamic movements while generalizing to tasks it has not encountered before.