SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model
Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive.
ProofPaper ↗
Key points
- 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.
- 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.
- SuperMimic removes compute as the bottleneck to scaling humanoid whole-body control, making it practical to measure motion-tracking scaling laws.
Sources (1)
- [1]SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World ModelarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 04:29 PM
Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive.
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.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0
- Oct 20, 2026Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds
- Oct 19, 2026NVIDIA Cosmos3-H-Dreams: Bringing Clinical Surgery to Life Through Generative Simulation
- Oct 19, 2026Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions
- Oct 19, 2026Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience
- Oct 19, 2026From Short Clips to Real-Time Streams: Identity-Aware Supervision and Context-Efficient Generation for Video Face Swapping