Humanoid Transformer

Humanoid Transformer is a scalable transformer architecture for whole-body humanoid control, proposed by researchers from Shanghai AI Lab, CUHK, SJTU, Zhejiang University, Peking University, Tsinghua University, and Galbot. It presents a principled scaling recipe for Behavior Foundation Models (BFMs) tailored to humanoid whole-body control. The architecture reformulates diverse humanoid control problems as a unified goal-conditioned RL objective of reproducing integrated whole-body trajectories from reference motions in the global frame, using PPO for on-policy training. It reduces MPKPE by 82% in global mode compared to baselines. Humanoid Transformer enables natural whole-body coordination, precise real-time responses to diverse control signals, and robust generalization across tasks and environments. It supports versatile humanoid behaviors including dexterous manipulation, natural and agile locomotion, and whole-body coordinated loco-manipulation under multiple control modes. The system has been validated in both simulation and real-world deployment on humanoid robots.