BlinkMimic: Training a Humanoid Motion Tracker in Minutes on a Single GPU
Abstract
Whole-body motion tracking lets a humanoid robot follow reference motions, but training a tracking policy for a library of motions has usually taken hours of GPU time. SuperMimic showed that the policy can instead be trained by backpropagating through a world model learned from parallel simulation, and InstantTrack added Muon updates, a residual policy around the reference and an averaged deployed policy. We build on both and present BlinkMimic, a recipe that makes this training faster. On a simulated Unitree G1, BlinkMimic trains one policy for all 40 LAFAN1 clips in 3.4 minutes on a single RTX 4090, measured until three consecutive checkpoints reach a normalized return of at least 0.82 with no failed clip (mean of three seeds). We split training time into the cost of an iteration, the iterations needed to fit the motions, and the iterations needed to remove the last failures, and assign each component, inherited or new, to one of these terms. The new components are a step size that decreases once training has progressed, resets directed by failures of the deployed policy under deployment conditions, and GPU graph capture of the whole iteration. A leave-one-out ablation with three seeds per component shows which term each component affects, and further measurements of the gradient, the update rule and the failure cases explain these effects. Code and per-seed results are available at https://github.com/fuyichen0725/blinkmimic.
Figures & tables
| Configuration | ms/it | Iterations to 0.82 | First 0.82 | 3 clean checkpoints | Tail failures | Final return |
|---|---|---|---|---|---|---|
| InstantTrack v1.2 (re-run) | 147 | 2408 / 2741 / 2492 | 5.6 / 6.3 / 5.8 (5.9) | 7.3 / 8.5 / 6.3 (7.4) | 80 / 126 / 113 | 0.856 |
| BlinkMimic (release) | 135 | 1292 / 1456 / 1169 | 2.89 / 3.28 / 2.64 (2.94) | 3.32 / 3.28 / 3.53 (3.38) | 2 / 6 / 0 | 0.861 |
| Removed | Term | It. to 0.82 | 3 clean (iteration) | Tail | Return |
|---|---|---|---|---|---|
| Full recipe | 1422 | 2100 / 2200 / 1900 | 1 / 8 / 0 | 0.860 | |
| Muon on the policy (Adam instead) | not by 6000 | – / – / – | 15 / 3 / 18 | 0.796 | |
| Muon on the world model | 1768 ( 24%) | 3000 / 3100 / 2100 ( 32%) | 24 / 22 / 7 | 0.843 | |
| Progress-triggered step size (constant ) | 1616 ( 14%) | 3000 / 1800 / 2300 ( 15%) | 20 / 3 / 6 | 0.856 | |
| Reference residual | 1751 ( 23%) | 1800 / 2700 / 2200 ( 8%) | 0 / 20 / 8 | 0.845 | |
| Semi-implicit world model and mirror | 1428 ( 0%) | 1800 / 2900 / 1900 ( 6%) | 0 / 15 / 13 | 0.854 |
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Metric | At the target | Iteration 3500 |
|---|---|---|
| Body position (cm) | 4.10 (4.02–4.14) | 3.45 (3.40–3.49) |
| Body orientation (deg) | 8.1 (7.9–8.3) | 6.9 (6.8–7.1) |
| Joint angle (deg) | 4.0 (4.0–4.1) | 4.2 (4.1–4.4) |
| Root drift (cm) | 93 (89–97) | 75 (73–77) |
| Foot slip (cm/s) | 7.1 (7.0–7.2) | 6.7 (6.6–6.7) |
| Contact agreement | 0.747 (0.744–0.749) | 0.764 (0.763–0.765) |
| Configuration | seeds | it. 0.82 | min | ms/it | strict it. |
|---|---|---|---|---|---|
| InstantTrack v1.0 (our run) | 2+1 | 4997 | 15.3 | 195 | 6167 (19.6) † |
| + semi-implicit WM, mirror | 3+1 | 4066 | 12 | 190 | – |
| + Muon (policy, WM) | 2 | 2598 | 7.6 | 194 | – |
| + reference residual, no warm-up | 1 | 1732 | 5.8 | 195 | 2750 (9.1) |
| + sampler/contact fixes (nb) | 3 | 1587 | 5.5 | 195 | 2750 / 2500 / – |
| + update graphs, allocation | 3 | 1549 | 4.9 | 172 ‡ | 2500 / 3000 / 3000 |
| Configuration | ms/it | Iterations to 0.82 | First 0.82 | 3 clean checkpoints | 6 clean checkpoints | Tail failures | Final return |
|---|---|---|---|---|---|---|---|
| No allocation (C) | 121 | 1439 / 1302 / 1479 | 3.2 / 3.0 / 3.2 (3.1) | 6.4 / 6.2 / 5.6 (6.1) | – / – / 5.6 (–) | 23 / 22 / 10 | 0.862 |
| Constant step size (A) | 142 | 1778 / 1679 / 1768 | 4.2 / 3.9 / 4.3 (4.1) | 6.6 / 6.7 / 5.1 (6.1) | 6.6 / 6.7 / 5.1 (6.1) | 9 / 14 / 5 | 0.862 |
| Progress-triggered step (B) | 141 | 1344 / 1331 / 1547 | 3.2 / 3.2 / 3.7 (3.3) | 4.8 / 5.7 / 4.1 (4.9) | 4.8 / 5.7 / 4.1 (4.9) | 2 / 10 / 2 | 0.861 |
| BlinkMimic (release) | 135 | 1292 / 1456 / 1169 | 2.89 / 3.28 / 2.64 (2.94) | 3.32 / 3.28 / 3.53 (3.38) | 5.36 / – / 3.53 (–) | 2 / 6 / 0 | 0.861 |
| Intervention | Effect | Reading |
|---|---|---|
| Batch , (teacher regression) | updates | noise is not limiting |
| Prioritized WM windows; prioritized imagination starts | within seed noise | reweighting changes variance only |
| Planning (CEM) teacher + behavior cloning, 4 variants | no strict-target speed-up; 1.6–1.7 wall | ; the student does not learn a margin |
| CEM data at failure points into the world model | cliff error to ; policy worse where taught | model accuracy at failures is not the limit |
| Critic slope supervision; critic variants | no gain; 2 slower per iteration | slope is not limiting |
| Mixed pathwise/likelihood-ratio estimator | offline – ; zeroth-order term 100–1600 noisier | pathwise variance does not grow exponentially |
| Experiment | Seeds | Result | Reading |
|---|---|---|---|
| Fit phase: signal side | |||
| Semi-implicit WM (on v1.0) | 3 | smaller Jacobian error | |
| + mirror augmentation (smir) | 3+1 | for both together (4066 vs. 4997) | symmetry prior; WM direction cosine 0.73 0.80 |
| WM Muon (alone / on smir+policy Muon) | 1 / 2 | / total | WM loss at 500 halved |
| Reference residual (on smir+Muon) | 1 | iterations, updates | RL-only gain (teacher unchanged) |
| No WM-only warm-up | 1 | iterations, wall to 0.82 | kept for simplicity |
Humanoid Motion Tracking
Humanoid motion tracking trains robots to reproduce human-like whole-body movements while remaining physically stable and feasible.
See all 91 papers