Driving · 2026-10-21
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
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
SuperMimic trains humanoid whole-body tracking using one learned differentiable world model trained from logged simulator transitions. The world model and goal-conditioned policy train together in a loop; backpropagation-through-time through short imagined rollouts provides analytic policy gradients. In 3 hours on one GPU, the authors report surpassing a SONIC baseline trained for 7 days on 128 GPUs, with >1000× fewer GPU-hours and fewer failed motions.
Why it matters
The authors argue SuperMimic removes compute bottlenecks to scaling humanoid whole-body control across many motions by amortizing shared dynamics in one differentiable world model. They report stable prediction error as motion coverage grows and demonstrate sim-to-real transfer to a Unitree G1. They also provide controlled comparisons showing faster, differentiable imagined rollouts and analytic gradients can reduce training cost.
Method
- Train a differentiable world model (dynamics, reward, observation, termination heads) from logged simulator transitions using k-step autoregressive replay windows (k=8).
- Jointly train a goal-conditioned policy and critic by backpropagating an imagined return through H=32-step frozen world-model rollouts (analytic gradients via BPTT).
- Use a Dyna-style loop: policy collects transitions into a replay buffer, the world model fits on replay windows, then the policy improves through imagined rollouts seeded from replay states.
Limitation
The authors report that on the filtered BONES-SEED 129,514-motion benchmark, SuperMimic fails on only 25 motions after 3 hours on one GPU, and fails on none after 9.8 hours on two GPUs.
Abstract (from arXiv)
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.
Related papers
- VioLA: Learning Generalist Humanoid Control Policies from Human Data
- BeJEPA: A Goal-Conditioned Humanoid Behavior World Model via Unsupervised RL
- Being-M0.7: A Latent World-Action Model for Humanoid Robots
- Humanoid World Action Model With Joint State--Action Generation
- GreenWAM: A World Foundation Model for Robotics
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-20 · Indexed by Pub-AI: 2026-10-21 UTC
SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model
Preprint posted: 2026-10-20 · Original record ↗
- arxiv
- 2610.28480
Explicit source versions (1)
- v1 · 2026-10-20
- Preprint available
Source date: 2026-10-20 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.