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

TL;DR

SuperMimic trains one differentiable world model on logged transitions from a standard rigid-body simulator and co-trains a motion-tracking policy against it, using backprop-through-time over short imagined rollouts to get analytic policy gradients. The result is a >1000× GPU-hour reduction vs a model-free RL motion-tracker baseline: 3 h on 1 GPU surpasses the baseline’s 7 days on 128 GPUs, with 25/129,514 benchmark-motion failures vs 179, and zero failures at 9.8 h on 2 GPUs. The world model’s prediction error stays flat from 1 to 264 hours of motion, and the policy transfers to a physical Unitree G1 across acrobatic motions — framing compute, not data or algorithms, as the current scaling bottleneck for humanoid whole-body tracking.

Key claims

Amortizing motion tracking across all motions with one differentiable world model, trained jointly with the tracking policy on logged rigid-body-simulator transitions, replaces per-motion specialists + distillation and brute-force PPO with analytic BPTT gradients from short imagined rollouts [Abstract].
3 h on 1 GPU surpasses the return of a SOTA motion-tracker baseline trained 7 days on 128 GPUs — a reported >1000× GPU-hour reduction [Abstract].
Benchmark failure count drops from 179 to 25 out of 129,514 motions at the compute-reduced setting; zero failures reached after 9.8 h on 2 GPUs [Abstract].
World-model prediction error stays stable as the motion corpus scales from 1 to 264 hours, and GPU-hours to a target return grow slowly over the same range [Abstract].
The trained policy transfers to a physical Unitree G1 across diverse and acrobatic motions [Abstract].

Method

SuperMimic couples two learned components inside one training loop. (1) A differentiable world model is fit from logged (state, action, next-state) transitions produced by a standard rigid-body simulator — the simulator remains the ground truth for data generation, but the world model is what the policy differentiates through. (2) A motion-tracking policy conditioned on a target reference clip is trained by backpropagating a tracking loss through short imagined rollouts under the world model, giving analytic policy gradients rather than model-free score-function estimators. The two are updated together so the world model continues to see on-policy transitions. The paper frames the imagined rollout as a short horizon with analytic gradient — a classical BPTT-through-learned-dynamics recipe scaled to a single generalist world model rather than per-motion specialists.

Results

Return surpassing the SOTA motion-tracker baseline in 3 h on 1 GPU vs 7 days on 128 GPUs, >1000× GPU-hour reduction [Abstract].
25 of 129,514 benchmark motions fail at that compute-reduced setting vs 179 for the baseline [Abstract].
Zero failures after 9.8 h on 2 GPUs [Abstract].
World-model prediction error stable from 1 h to 264 h of motion data; compute-to-target-return grows slowly over the same range [Abstract].
Sim-to-real transfer on a physical Unitree G1 across diverse and acrobatic motions, no per-motion specialization [Abstract].

Why it’s interesting

SuperMimic is the sharpest filed instance of the differentiable-world-model + BPTT policy-learning recipe applied at humanoid whole-body-tracking scale, and it stakes a position directly on the central debate tracked by World Foundation Models: the WFM here is not a video-rollout prior (Cosmos 3, Genie 3) and not a latent-predictive JEPA (LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics) — it is a learned dynamics approximator over sim transitions whose load-bearing property is differentiability, used to produce analytic policy gradients. Complement to Scaling Behavior Foundation Model for Humanoid Robots ScaleBFM’s PPO-on-retargeted-mocap recipe (model-free; retargeting-source diversity as the lever) by replacing the model-free RL inner loop with a model-based analytic-gradient inner loop at reportedly >1000× lower GPU-hour cost. The physical-G1 transfer plus 264-hour-corpus flat prediction error is also a counterpoint to Jiahua Dong — 30,000-hour egocentric world model learns hands cleanly, not objects; neither 100× data nor 4× model closes the gap Jiahua Dong’s “~10⁴× more compute needed to close the object-dynamics gap” measurement on 30k-hour pure-pixel human-video WFMs — SuperMimic argues that for whole-body tracking the right WFM is the one that lives in simulator state space, where contacts and forces are observed, rather than in pixel space where Dong showed they are not.

On Human-to-Robot Retargeting, SuperMimic sidesteps the interface-location debate entirely at training time: there is no human-to-robot retargeting step inside the SuperMimic loop (the input is simulator transitions, not retargeted human mocap). Retargeting lives upstream in whatever pipeline produces the 129,514 reference motions — but the compute reduction here means that the retargeted corpus can grow by orders of magnitude without the policy-training cost scaling with it, which is a different lever on the retargeting-quality-vs-retargeting-diversity trade-off than Compute-Optimal Scaling Laws for Human Motion Generation Cartwheel’s data-side scaling law.