Market
SuperMimic paper reports training a Unitree G1 motion controller in about three hours on one GPU, while “minutes” claims come only from unofficial open-source repos
SuperMimic reports training a Unitree G1 motion controller in about 3 hours on one H100. “Minutes” appears only in unofficial repos on small clip sets.
The SuperMimic paper (arXiv 2610.28480) reports training a motion-tracking controller for the Unitree G1 humanoid in about three hours on a single H100 GPU, against a baseline it describes as taking seven days on 128 GPUs. The “a few minutes on one GPU” version of the claim is not in the paper: it comes from unofficial open-source repositories that train on small sets of about 40 clips. Both are authors’ or developers’ own reports, and we have not reproduced either.
The essentials
- The paper’s claim: about 3 hours on one H100, versus 7 days on 128 GPUs for the SONIC baseline, per the alphaXiv page.
- The gap: the authors describe more than 1,000 times fewer GPU-hours.
- The quality figure: 25 failures versus 179 on 129,514 motions, as reported by the authors.
- The “minutes” claim: it appears only in unofficial repositories, trained on about 40 LAFAN1 clips.
Hours in the paper, minutes in the repos
The paper’s own number is about three hours on one H100, which is already a large reduction from the baseline it cites. The faster “minutes” figure belongs to separate community repositories such as BlinkMimic, which train on a few dozen clips. Training on 40 clips and on more than 129,000 motions are different workloads, so the two numbers should not be set side by side as one result.
Caveats
The comparison with SONIC is the authors’ own, and a G1 demonstration shows the controller on a real robot rather than a broad benchmark of real-world reliability. Counting failures on simulated motions is not the same as success on hardware. We read the alphaXiv summary page directly and did not run any code.
What is not established
- Independent reproduction of the three-hour result: none found.
- Real-robot reliability beyond the G1 demonstration: not shown in the sources read.
- Official support for the “minutes” claim: none; it comes from unofficial repositories.
- Whether the repos’ speed holds on large motion sets: not tested in the sources read.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Paper | arXiv 2610.28480, SuperMimic | arXiv |
| Reported training cost | About 3 hours on one H100 | alphaXiv |
| Baseline compared | SONIC, 7 days on 128 GPUs, as described by the authors | alphaXiv |
| Failures | 25 versus 179 on 129,514 motions, per the authors | alphaXiv |
| “Minutes” claims | Unofficial repositories, about 40 LAFAN1 clips | GitHub |
FAQ
Does the paper say a controller trains in minutes?
No. The paper reports about three hours on one H100; “minutes” comes from unofficial repositories on small clip sets.
Is the result independently reproduced?
We found no independent reproduction of the paper’s numbers.
Is it tested on a real robot?
The paper includes a Unitree G1 demonstration, but the sources we read do not show broad real-world reliability data.
Method and limits
Checked on November 16, 2026. We read the alphaXiv page directly; the arXiv PDF and the GitHub repositories were seen through search excerpts only. We ran no code.
Sources
- alphaXiv: SuperMimic (arXiv 2610.28480) (verified 2026-11-08)
- arXiv: SuperMimic paper 2610.28480 (verified 2026-11-08)
- GitHub: BlinkMimic open-source repository (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy