# SuperMimic

> Source: https://aiwiki.ai/wiki/supermimic
> Updated: 2026-10-22
> Categories: AI Research, Reinforcement Learning, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "SuperMimic." aiwiki.ai, 22 Oct 2026. https://aiwiki.ai/wiki/supermimic
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

SuperMimic is a [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) method for training a [humanoid robot](https://aiwiki.ai/wiki/humanoid_robot) controller to follow human motions. Its Berkeley researchers use a differentiable [world model](https://aiwiki.ai/wiki/world_model) to reduce the training compute needed for motion tracking.[1][2]

## Training method

The controller collects transitions in MuJoCo Warp, a GPU implementation of [MuJoCo](https://aiwiki.ai/wiki/mujoco). A learned dynamics model fits those transitions, and the controller improves through imagined trajectories. Gradients pass through that model rather than the simulator. Data collection, model fitting and policy improvement alternate within one training loop.[2]

## Reported comparison

The authors compared their system with SONIC's released checkpoint, rather than retraining SONIC under identical conditions.[2][3]

| Condition | SONIC | SuperMimic |
|---|---|---|
| Training hardware | 128 NVIDIA L40 GPUs | One NVIDIA H100 GPU |
| Training time | 168 hours | Three hours to exceed SONIC's return |
| Motion data | 611 hours | 264.56 hours of BONES-SEED |
| Failed benchmark motions | 179 of 129,514 | 25 of 129,514 |

These are the researchers' reported conditions.[2][3] Different GPUs and datasets prevent treating the GPU-hour comparison as a hardware-normalized, universal speedup. SuperMimic's training budget includes simulator collection, dynamics learning and policy updates, but excludes subsequent checkpoint evaluation.[3]

In a separate controlled test on 40 LAFAN1 clips, SuperMimic reached normalized return 0.82 in 1.2 hours on one H100. The authors' matched PPO baseline did not reach it within 18 hours on one, four or eight GPUs.[2]

## Evaluation and physical demonstrations

After 9.8 hours on two H100 GPUs, the authors reported no failures across 129,514 clips. These clips were training motions, so the result measures coverage rather than unseen-motion generalization. Evaluation used deterministic simulation; the termination criterion also differed from SONIC's original criterion.[2][3]

The paper demonstrates a single controller performing a skipping jump, side roll and two flips on a physical [Unitree G1](https://aiwiki.ai/wiki/unitree_g1). This is narrower evidence than the large simulated motion test. Training used simulator data and flat ground, without terrain or object interaction.[2][3]

## Publication and materials

The paper was submitted on October 20, 2026. Its arXiv record and Berkeley laboratory listing identify acceptance for an oral presentation at CoRL 2026. Acceptance does not establish independent reproduction. The project website includes demonstrations and an interactive viewer.[1][4]

## References

1. Liao, Qiayuan, et al. [SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model](https://arxiv.org/abs/2610.28480). arXiv:2610.28480v1, October 20, 2026.
2. Liao, Qiayuan, et al. [Full paper](https://arxiv.org/pdf/2610.28480). 2026. Sections 3-6 and appendices.
3. SuperMimic authors. [Project website](https://supermimic.github.io/). Accessed October 23, 2026.
4. UC Berkeley Hybrid Robotics Group. [Publications](https://hybrid-robotics.berkeley.edu/publications/). CoRL 2026 entry, accessed October 23, 2026.