Newest papers

SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model

Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel, Koushil Sreenath · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28480 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

El seguimiento práctico de los grandes conjuntos de datos de movimiento humano se ha convertido en la ruta dominante para el control general, expresivo de todo el cuerpo humanoide, pero la escala en muchos movimientos es costosa. Las tuberías actuales utilizan el aprendizaje de reforzamiento sin modelo, ya sea mediante la fusión de especialistas de movimiento por movimiento a través de la destilación de pérdidas, o mediante la fuerza bruta de una sola política con la computación a escala de clúster. SuperMimic amortiza el aprendizaje en todos los movimientos con un modelo mundial diferenciable de su dinámica compartida, ajustado a las transiciones de movimiento logradas de un simulador de cuerpo rígido estándar. El modelo mundial y el tren de seguimiento de la política se unen en un solo ciclo, y la propagación a través del tiempo

Machine translation (Meta M2M100); the English original follows.

Abstract

Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive. Current pipelines use model-free reinforcement learning, either merging per-motion specialists through lossy distillation, or brute-forcing a single policy with cluster-scale compute. SuperMimic amortizes learning across all motions with one differentiable world model of their shared dynamics, fit from logged transitions of a standard rigid-body simulator. The world model and the tracking policy train together in one loop, and backpropagation-through-time over short imagined rollouts gives the policy analytic gradients. Fast imagined rollouts and analytic gradients make policy learning efficient. In 3 hours on a single GPU, SuperMimic surpasses the return of a state-of-the-art motion tracker baseline trained for 7 days on 128 GPUs. It reduces GPU-hours by more than $1000\times$ and fails on only 25 of 129,514 benchmark motions, versus 179 failures for the baseline. After 9.8 hours on two GPUs, SuperMimic achieves zero failures. From 1 to 264 hours of motion, the world model's prediction error stays stable, and GPU-hours to a target return grow slowly. Our policy transfers to a physical Unitree G1 across diverse and acrobatic motions. SuperMimic removes compute as the bottleneck to scaling humanoid whole-body control, making it practical to measure motion-tracking scaling laws.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Reinforcement learning, Computer science, Scaling, Bottleneck, Benchmark (surveying), Lossy compression

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange