Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

SuperMimic

UC Berkeley's training code for a humanoid motion tracker learned through one differentiable world model, on one GPU in hours.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.8/5

SuperMimic turns humanoid motion tracking from a cluster job into an afternoon on one GPU, but today it is research code you have to retrain yourself.

What is SuperMimic?

SuperMimic is the code behind a UC Berkeley paper, accepted as an oral at CoRL 2026, from Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel and Koushil Sreenath. It trains one policy that makes a humanoid robot copy whole-body human motion, from walking to flips and side rolls. Instead of running reinforcement learning straight in the simulator, it fits a single learned world model to transitions logged from MuJoCo Warp and trains the policy through that model with analytic gradients over short imagined rollouts of 32 steps.

The repository holds Python training and playback scripts for two task IDs on the Unitree G1, one using generalised and one using maximal coordinates. It is built on mjlab and needs HybridRobotics' mjlab_research fork cloned beside it. The project page carries a browser demo that runs the policy on a chosen motion.

The repository was created on 10 November 2026, is Apache 2.0-licensed and has about 1,893 GitHub stars and 149 forks. It has no tagged releases; the last commit landed on 19 November 2026.

Is SuperMimic worth it?

The headline claim is about compute, and the project page backs it with its own accounting. On a benchmark of 129,514 motions, SuperMimic beats the normalised return of SONIC's released checkpoint (0.90) after 3 hours on one NVIDIA H100. SONIC was trained for 7 days on 128 NVIDIA L40 GPUs, about 21,000 GPU-hours, so the gap is more than 1000 times in GPU-hours. SuperMimic also trained on less data: 264.56 hours of BONES-SEED, the public subset of SONIC's 611 hours.

The failure counts are the more telling figure. SONIC's checkpoint fails 179 motions, including all 44 flips and all 22 side rolls. After 3 hours SuperMimic fails 25, among them 1 flip, 2 side rolls and 8 lying or standing-up motions; after 9.8 hours on two GPUs the authors report none. The page is careful to say that the benchmark motions are training motions for both methods, so these numbers measure how much of the library a policy covers, not how it handles motion it has never seen.

The method does not need a differentiable simulator. The world model learns from an ordinary contact simulator, updates every iteration alongside the policy, and a critic estimates the return beyond the 32-step imagined horizon. The authors say one policy runs zero-shot on a physical Unitree G1 in every real-robot clip they show, with no per-motion tuning. That rests on videos and the paper; how robust it is outside those clips should be checked in the paper's real-robot section before you plan hardware time around it.

The catch: the repository is training-only. An open issue notes there are no configs for the paper's runs, no data download path and no released checkpoint, and another reports that the README's training command fails for outside accounts because the motion data sits in a Weights & Biases registry they cannot reach. Playback also reads runs from W&B. The project is ten days old, depends on a 47-star research fork of mjlab, and pins Python 3.13.

Our take: for a humanoid lab, SuperMimic is the result to reproduce this quarter, because it moves whole-body tracking from a budget line to a single workstation. For anyone who wants a policy to drop onto a G1 tomorrow, it is a paper to read while the code catches up.

Where it earns its place

  • A university humanoid lab with one or two H100s

    Train a general motion tracker for a Unitree G1 without renting a cluster

    Clone supermimic and mjlab_research side by side, prepare your own motion set until the data path is public, and train the SuperMimic-GenCoord-G1 task. The paper's run beat its baseline in 3 hours on one H100; plan extra time to rebuild the paper's configs.

  • A robotics startup evaluating world-model training against PPO

    Decide whether model-based training cuts their GPU bill

    Run both coordinate variants on a slice of your own motions and compare GPU-hours to a target return. The authors report about 64 times fewer imagined steps than PPO on the same world model to reach PPO's best training return.

Before you commit

Setup
Install uv, clone supermimic and HybridRobotics/mjlab_research as sibling folders, then run the train script with a task ID on a chosen GPU. Python 3.13 is required.
Integrations
Built on mjlab and MuJoCo Warp; logs and loads runs through Weights & Biases; ships Unitree G1 tasks only.
Data and privacy
Runs on your own machines. The training motions come from BONES-SEED; the README's motion artefact sits in a W&B registry that outside accounts report they cannot reach.
Cost at scale
Free code. The cost is GPU time: 3 GPU-hours on an H100 to pass the baseline, 19.6 GPU-hours (9.8 hours on two GPUs) for zero failed motions in the paper.
Limits
No checkpoints, no evaluation harness and no tagged releases; G1 only; real-robot robustness is shown in clips and should be checked in the paper.

What’s good

  • Beats a 128-GPU baseline after 3 hours on one H100
  • 25 failed motions of 129,514 against 179, and none after 9.8 hours on two GPUs
  • Handles flips and side rolls that the SONIC checkpoint fails
  • Apache 2.0, with a browser demo of the trained policy

What’s not

  • No released checkpoint or configs for the paper's runs
  • Motion data path in the README fails for outside accounts
  • Benchmark motions are training motions, so it measures coverage, not generalisation
  • Depends on a young research fork of mjlab and Weights & Biases

Best for

Robotics researchers with an NVIDIA GPU who want to reproduce or extend whole-body humanoid tracking without a GPU cluster.

SuperMimic alternatives

SuperMimic: common questions

Is SuperMimic worth it?
SuperMimic turns humanoid motion tracking from a cluster job into an afternoon on one GPU, but today it is research code you have to retrain yourself. We rate it 3.8 out of 5.
How much does SuperMimic cost?
Free and open source under Apache 2.0. You need your own NVIDIA GPU; the paper's headline run used one H100 for 3 hours.
What are the downsides of SuperMimic?
No released checkpoint or configs for the paper's runs; motion data path in the README fails for outside accounts; benchmark motions are training motions, so it measures coverage, not generalisation; depends on a young research fork of mjlab and Weights & Biases.
Who is SuperMimic best for?
Robotics researchers with an NVIDIA GPU who want to reproduce or extend whole-body humanoid tracking without a GPU cluster.
How much does SuperMimic cost for a team or in production?
Free code. The cost is GPU time: 3 GPU-hours on an H100 to pass the baseline, 19.6 GPU-hours (9.8 hours on two GPUs) for zero failed motions in the paper.
Is SuperMimic safe for company data?
Runs on your own machines. The training motions come from BONES-SEED; the README's motion artefact sits in a W&B registry that outside accounts report they cannot reach.
What are the best SuperMimic alternatives?
The closest alternatives we have reviewed are Eden.

Answers checked Nov 20, 2026.