Scaling Behavior Foundation Model for Humanoid Robots
The paper reports large humanoid-control gains by coordinating motion-tracking training, rollout scale, motion diversity, and a Humanoid Transformer architecture.
The authors frame varied humanoid tasks as reproducing whole-body behavior in a global frame. They say performance depends on pairing enough on-policy rollouts with diverse reference motions, rather than scaling one piece alone. In simulation and real-world deployment, their model cut test MPKPE by more than 10% in local mode and 82% in global mode versus existing humanoid controllers. ArXiv · AI/CL/LG's note
The authors frame varied humanoid tasks as reproducing whole-body behavior in a global frame. They say performance depends on pairing enough on-policy rollouts with diverse reference motions, rather than scaling one piece alone. In simulation and real-world deployment, their model cut test MPKPE by more than 10% in local mode and 82% in global mode versus existing humanoid controllers. ArXiv · AI/CL/LG's note
score 5