Recoleta
GitHub
Source note

VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon

Authors
Yi Pan, Miao Pan, Qi Lu, Jiaming Huang, Man Zhang, Siteng Huang, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot PolicyAdaptive Action HorizonInference Time CorrectionLatent Dynamics
Open arXivMarkdown source

Summary

VLA-Corrector adds inference-time monitoring and correction to action-chunked vision-language-action policies. It raises task success while often reducing policy calls by cutting off stale action chunks and guiding the next replan.

Problem

  • Action-chunked VLA policies execute several predicted actions open-loop, so fresh camera observations do not affect control until the chunk ends.
  • Longer horizons reduce policy-call cost, but they create more time for contact errors, pose drift, and slippage to compound into task failure.
  • Fixed horizons are hard to tune because the best value changes with task difficulty, dynamics, and sim-to-real mismatch.

Approach

  • A frozen VLA backbone still generates action chunks; VLA-Corrector adds a separate 40M latent dynamics MLP trained on demonstration trajectories.
  • The Latent-space Vision Monitor predicts the expected visual latent change for the executed action and compares it with the actual latent change from new camera frames.
  • A sliding-window median and MAD threshold detect persistent mismatch; when drift persists for several steps, the system discards the rest of the current action queue.
  • After an interrupt, Online Gradient Guidance modifies the flow-matching denoising step so the next action chunk points toward the latent correction direction.
  • This creates an adaptive horizon: chunks run long when visual dynamics stay on track and shorten when execution drifts.

Results

  • On MetaWorld, pi0.5 average success rises from 48.70% to 64.35% (+15.65 points); its Very Hard split rises from 41.0% to 65.0% (+24.0 points).
  • On MetaWorld, SmolVLA improves from 61.90% to 66.65% (+4.75 points), and X-VLA improves from 55.55% to 59.60% (+4.05 points).
  • On LIBERO, few-shot pi0.5 plus VLA-Corrector improves average success from 94.00% to 97.80% (+3.80 points), above the fully fine-tuned pi0.5 baseline at 96.95%.
  • For pi0.5 at horizon 50, success improves from 48.72% to 58.70%, average policy calls drop from 5.15 to 4.98, and success-per-call efficiency gains 24.6%.
  • For SmolVLA at horizon 10, success improves from 61.90% to 73.00%, policy calls drop from 19.27 to 15.64, and success-per-call efficiency gains 45.3%.
  • Corrector training uses limited data well: on MetaWorld at horizon 50, using the full post-holdout corrector training split improves average success from 48.72% to 54.32%, while r=0.6 reaches 52.20%.

Link