Driving · 2026-11-13

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

Fuyi Chen

arXiv:2611.16422PDFCode

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

BlinkMimic is a single-GPU training recipe for whole-body humanoid motion tracking, building on SuperMimic and InstantTrack. On simulated Unitree G1, the authors report one policy for all 40 LAFAN1 clips reaches a target of mean normalized return ≥0.82 with zero failed clips over three consecutive checkpoints in 3.4 minutes on an RTX 4090. The paper attributes speedups to Muon updates, a progress-triggered step-size schedule, residual actions, deployment-failure probing, and CUDA graph capture of whole iterations.

Why it matters

The authors address that training a tracker over a library of motions has usually taken hours to days on multiple GPUs. BlinkMimic aims to make time-to-target on stated hardware a controllable design objective by separating training time into a “fit phase” (getting mean return to 0.82) and a “tail phase” (eliminating the last failures). They provide an explicit recipe, ablations, and measures of how each ingredient affects those phases and iteration cost.

Method

  • Trains one policy for 40 LAFAN1 clips on simulated Unitree G1 using BPTT through a learned world model, with an EMA deployed policy and MujoCo-Warp at 50 Hz.
  • Splits training time into fit vs tail phases and assigns ingredients accordingly: Muon updates (policy and world model), a progress-triggered step-size schedule, residual policy around reference, and deployment-condition failure probing with reset reallocation.
  • Reduces per-iteration cost by capturing the whole iteration (including simulation/episode resets) as CUDA graphs and using a smaller imagination batch size.

Limitation

The deployed-policy evaluation is deterministic and does not use domain randomization or noise; the paper does not state results for real robots or other environments beyond the simulated Unitree G1 setting described.

Abstract (from arXiv)

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.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-11-12 · Indexed by Pub-AI: 2026-11-13 UTC

preprint

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

Fuyi Chen

Preprint posted: 2026-11-12 · Original record ↗

arxiv
2611.16422
Explicit source versions (1)
  • v1 · 2026-11-12
  1. Preprint available

    Source date: 2026-11-12 · First detected: 2026-11-13 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

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.