InstantTrack
Fifteen minutes.
One GPU.
Whole-body motion tracking for the Unitree G1, trained on every LAFAN1 clip.
1 GPU · 15 min to return 0.82 with 0 of 40 clips failed.2.2× faster than SuperMimic + EMA on the same machine.
CodeGet the highlights.
The same 15 minutes of training.
40 motions. One policy.
After 15 minutes on one GPU.
After 1.5 hours both track every clip.
InstantTrack stays closer to the reference.
Speed
Fifteen minutes.
Not thirty-three.
SuperMimic with an EMA policy needs 0.55 h to reach return 0.82 with no failed clip. InstantTrack needs 0.25 h, on the same GPU, simulator and evaluator.
Normalized return against training time
40 LAFAN1 clips · mean of 3 seeds · 1 RTX 5090
Line chart of evaluation normalized return against training hours, from 0 to 0.75 h. InstantTrack crosses the 0.82 target (dashed) near 0.2 h and holds it from 0.25 h; SuperMimic plus EMA holds it from 0.55 h. Left and right arrow keys step through the checkpoints.
RatioInstantTrackSuperMimic + EMA
2×more training iterations per second
InstantTrack6.3 it/s
SuperMimic + EMA3.1 it/s
6×less time than SuperMimic without EMA
InstantTrack0.25 h
SuperMimic1.50 h
How it works
One loop. Rebuilt for the GPU.
InstantTrack keeps SuperMimic’s loop: one world model, analytic policy gradients through it. It makes every step of the loop cheaper.
Failed clips
Twelve. Ten. None.
After 15 minutes of training, SuperMimic + EMA still fails 9 to 12 of the 40 clips and SuperMimic fails 10, most of them falls and get-ups. InstantTrack fails none, in every seed.3
| Group (clips) | SuperMimic | SuperMimic + EMA | InstantTrack |
|---|---|---|---|
| Fall and get up (6) | 5 | 5–6 | none |
| Dance (8) | 2 | 2 | none |
| Walk (12) | 2 | 1–2 | none |
| Jumps (3) | 1 | 1 | none |
| Fight (3) | none | 0–1 | none |
| Fight and sports (2) | none | none | none |
| Run and sprint (6) | none | none | none |
| Total (40) | 10 | 9–12 | none |
Ranges are over 3 seeds; SuperMimic has one seed. The clips are training clips for every method, so the counts measure coverage, not generalization.
What changed.
Rotations that the compiler can fuse
315 → 209 msper training iteration. The 3×3 rotation products in kinematics, reward and world model are written element-wise, and reward, observation and goal are compiled. Kernel launches per iteration drop from 25.3k to 15.2k.4
A right-sized world model
+28%iterations per second. For 40 clips, a 1024³ world model learns as much per iteration as the paper’s 2048³.
An averaged policy
2.7×on SuperMimic alone. Polyak-averaged policy weights take SuperMimic from 1.50 h to 0.55 h, so we count the EMA as part of the baseline.
Resets aimed at failures
0.27 → 0.22 hmean time to the target. 30% of resets start one second before where episodes recently failed, so hard clips stop flickering between checkpoints.
A contact reward in imagination
0.74 → 0.81foot-contact agreement with the reference. The feet are rewarded for matching the reference contact state and for not sliding while it is down. Evaluation keeps SuperMimic’s reward.
Side by side.
SuperMimic + EMA
our reimplementation
InstantTrack
main configuration
Training method
Backpropagation through one learned world model
The same, with a GPU-native training loop
Hardware
1NVIDIA RTX 5090
1NVIDIA RTX 5090
Time to the target
0.55 h0.50 / 0.54 / 0.60 h over 3 seeds
0.25 h0.25 / 0.25 / 0.25 h over 3 seeds
Failed clips at 0.25 h
9–12of 40
noneof 40
Body position error at 0.75 h
4.60 cm14 tracked bodies
3.53 cm23% lower
Body orientation error at 0.75 h
11.1°
7.5°32% lower
Foot-contact agreement at 0.75 h
0.72
0.81
Foot slip in contact at 0.75 h
8.00 cm/sthe reference itself slips 10.2
6.62 cm/s17% lower
Global root drift at 0.75 h
54 cm
61 cmlarger: the feet do not slide with the reference
Questions? Answers.
Is this the official SuperMimic code?
No. No official SuperMimic code is released. Our SuperMimic is the mjlab port of supermimic_vibed, an independent reimplementation of the paper. The paper reports about 1.2 h to return 0.82 on an H100; we measure 1.50 h on an RTX 5090, under a stricter criterion. Every speed-up on this page compares InstantTrack with our SuperMimic on the same machine, not with the paper’s numbers.
Why compare with SuperMimic + EMA?
A Polyak-averaged policy alone takes SuperMimic from 1.50 h to 0.55 h, and any SuperMimic implementation can adopt it. So the fair baseline is SuperMimic + EMA, and the headline is 2.2×, not 6×.
What do the 15 minutes include?
The full training loop on an otherwise idle GPU: simulation, world-model fitting, policy and critic updates, torch.compile warm-up and checkpoint writing. Checkpoint evaluation runs afterwards and is not counted.
What counts as reaching the target?
Normalized return ≥ 0.82 with 0 of 40 clips failed, on 3 checkpoints in a row, so a single lucky checkpoint does not count. Checkpoints come every 3 minutes, so times are quantized to 0.05 h.
Where is InstantTrack worse?
With the contact reward, normalized return at 1.5 h is 0.877, against 0.880 for SuperMimic + EMA, and global root drift is larger. Without the contact reward, InstantTrack is better on every metric we measure, at 0.890 return.
What is not covered?
Only the Unitree G1 in simulation, on LAFAN1’s 40 clips. There are no evaluation-time pushes or randomization, no larger motion library and no real-robot experiment. Plain SuperMimic has one seed; the other methods have three.
What did not help?
A smaller policy, demonstration replay, policy batch 2048, EMA 0.999, lr 6e-4, horizons 16 and 24, a foot-contact world-model loss, anchored imagination, more critic epochs, UTD 0.5, CUDA graphs, world-model window 4 and a contact-mode world model. Each is in the README with its run logs.
- Normalized return is a clip’s return divided by the maximum possible over the whole clip, averaged over the 40 clips, from 0 to 1. All 40 clips play once from their first frame, with friction 1.0, no randomization or noise, and a deterministic policy.
- Time to return ≥ 0.82 with 0 of 40 failed clips, held for 3 consecutive checkpoints, mean of 3 seeds: InstantTrack 0.25 / 0.25 / 0.25 h, SuperMimic + EMA 0.50 / 0.54 / 0.60 h, SuperMimic 1.50 h (1 seed). Training wall-clock on one RTX 5090 with 4096 parallel environments; evaluation is not counted.
- Checkpoint nearest to 0.25 h of each run. A clip fails if the rollout triggers SuperMimic’s training termination (root height or tilt error).
- Measured under torch.profiler on the configuration before the smaller world model. Training speed rises from 3.6 to 5.0 iterations per second with these two changes.
Open source · MIT
InstantTrack: Whole-Body Motion Tracking for the Unitree G1 in 15 Minutes on One GPU
Built on SuperMimic (Liao et al., 2026), supermimic_vibed (Chong Zhang), InstantMimic (Choi, Leem, Won, 2026), mjlab, MuJoCo Warp and LAFAN1 (Ubisoft La Forge).
BibTeX
@misc{instanttrack2026,
title = {{InstantTrack}: Whole-Body Motion Tracking for the
Unitree G1 in 15 Minutes on One GPU},
author = {Huy-Quang-Dao},
year = {2026},
url = {https://github.com/Huy-Quang-Dao/instantTrack}
}
Please also cite SuperMimic, InstantMimic, mjlab and LAFAN1. InstantTrack is independent work, not affiliated with or endorsed by the SuperMimic authors. Videos show motions derived from LAFAN1 (CC BY-NC-ND 4.0), for non-commercial research and illustration only. Page design adapted from the SuperMimic project page.