Paper Describes Transformer-Based Scaling Approach for Humanoid Control

The authors report improved control fidelity and task generalization in simulation and real-world deployment, using a motion-tracking training approach and a Humanoid Transformer architecture.

A research team including Weishuai Zeng has described a behavior foundation model approach for humanoid robot control aimed at improving motion-tracking fidelity and task generalization.

The paper’s scaling approach combines three components: a motion-tracking learning paradigm that frames control problems as reproducing integrated whole-body behavior in a global frame; coordination between on-policy rollout quantity and reference-motion diversity; and a transformer-based policy architecture called Humanoid Transformer.

The authors argue that humanoid control requires whole-body coordination, real-time responses to control signals, and the ability to generalize across different environmental contexts. Their motion-tracking formulation is designed to represent diverse humanoid control problems through the reproduction of whole-body behaviors rather than as separate tasks.

The paper reports substantial performance gains in control fidelity and task generalization through experiments in simulation and real-world deployment. It uses Mean Per-Keypoint Position Error, or MPKPE, as a measure of motion-tracking accuracy.

Humanoid Transformer is intended to provide an expressive architecture that can scale with behavioral data and support the emergence of structured behavioral representations. The authors also emphasize the relationship between data collected through on-policy rollouts and the diversity of reference motions used during training.

The supplied abstract does not specify the physical humanoid platform, benchmark tasks, baseline controllers, or the particular real-world deployment setup. Those details would be important for assessing how the reported results translate to other hardware, motion sets, and operating conditions.