← Paper Digest

BlinkMimic: Training a Humanoid Motion Tracker in Minutes on a Single GPU

Fuyi Chen

November 2026

TL;DR A training recipe that fits one whole-body tracking policy for all 40 LAFAN1 motion clips on a simulated Unitree G1 in 3.4 minutes on a single RTX 4090, against 7.4 minutes for the InstantTrack baseline re-run on the same GPU.

What they do

Training a humanoid whole-body motion tracker for a library of motions has usually taken hours of GPU time. BlinkMimic builds on SuperMimic (backpropagating through a learned world model instead of policy gradients) and InstantTrack (Muon updates, a residual policy, an averaged deployed policy) and asks how far this training can be sped up. It splits 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 addresses each term.

How it works

  • Training loop. 4096 simulated Unitree G1s in MuJoCo-Warp collect data, a world model is fit to replay windows, and the policy follows the gradient backpropagated through 32-step imagined rollouts. An exponential moving average of the policy is deployed.
  • Fit phase. Muon updates for both the policy and the world model, a policy that outputs a correction to the reference joint targets, and a step size that starts at 4x the base rate and falls once the training return passes 0.75.
  • Tail phase. Every 125 iterations a separate simulator sweeps the whole library with the deployed policy under evaluation settings. The resulting failure map receives 15% of episode resets.
  • Iteration cost. The whole iteration, including physics and resets, is captured as CUDA graphs, and the imagination batch is halved to 2048.

Results

  • Reaches the target (three consecutive checkpoints with mean return at least 0.82 and no failed rollout) in 3.32, 3.28 and 3.53 minutes for three seeds. InstantTrack, re-run on the same GPU, takes 7.4 minutes on average.
  • At the target, the RMS position error of the 14 tracked bodies is 4.1 cm and the mean joint-angle error 4.0°.
  • Ablation: with Adam instead of Muon on the policy, the return stays at 0.81 up to iteration 6000. Removing Muon on the world model, the step-size schedule or the residual lengthens the fit phase by 24%, 14% and 23%.
  • Failed rollouts over checkpoints 2000-3500: 2 / 6 / 0 per seed, against 80 / 126 / 113 for InstantTrack.

Limitations

One robot, one motion library and one GPU type, with no real-robot deployment. A checkpoint that meets the target can fail later: in one seed a clip fails in all four replicas at iteration 3100. Three seeds cannot confirm the effect of the failure probe in the final recipe. The asynchronous probe can crash training when another process shares the GPU.

Figures

From Figure 1: One policy for 40 LAFAN1 clips in 3.4 minutes on one RTX 4090. (a) The deployed policy (solid) trackin...
From Figure 1: One policy for 40 LAFAN1 clips in 3.4 minutes on one RTX 4090. (a) The deployed policy (solid) tracking the reference (translucent) over the most dynamic seconds of three clips. (b) Training time to the first mean return of 0.82 (light) and to three consecutive clean checkpoints (dark), against InstantTrack and an estimate for SuperMimic. (c) Mean return over 40 clips against training time. (d) Where the 3.4 minutes go, and the cost of one iteration by stage.
From Figure 2: One iteration of BlinkMimic. The real simulator (blue) feeds the learned world model and critic (orang...
From Figure 2: One iteration of BlinkMimic. The real simulator (blue) feeds the learned world model and critic (orange), and the policy improves through the imagined gradient. Numbers mark the ingredients: (1) Muon updates for the policy and the world model; (2) a step size that falls on progress; (3) the reference residual; (4) the semi-implicit, mirror-augmented world model; (5) the deployment-condition probe and reset allocation; (6) the imagination batch and the whole iteration captured as GPU graphs.
From Figure 5: The tail is a practice problem. (a) Failed rollouts per clip and checkpoint over iterations 1500-3500,...
From Figure 5: The tail is a practice problem. (a) Failed rollouts per clip and checkpoint over iterations 1500-3500, without and with reset allocation. (b) Share of a hard window in the reset distribution, the visited states and the imagination starts. (c) Late failure rate of the hardest clip trained alone, in the 40-clip mix, and in the mix with 10% of resets forced onto the window before its hard spot.
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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.