AIAny
Icon for item

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

Detects when an action-chunked VLA policy drifts from expected visual dynamics and triggers lightweight corrective replanning via a latent-space vision monitor and online gradient guidance; creates an event-driven adaptive action horizon without retraining the backbone.

Introduction

Open-loop action chunking reduces policy-call frequency but creates blind spots where small perturbations can rapidly compound into failures in contact-rich tasks. VLA-Corrector reframes this trade-off by adding a lightweight, event-driven inference layer that watches latent visual evolution and interrupts stale action chunks only when persistent drift is detected, restoring closed-loop reactivity without modifying the base policy.

Key Findings
  • Latent-space Vision Monitor (LVM): continuously compares predicted vs. observed visual feature evolution to detect persistent deviation, enabling early truncation of unreliable action chunks — so what: preserves reliable long-horizon execution while catching divergence early.
  • Online Gradient Guidance (OGG): a lightweight corrective replanning step that optimizes remaining actions after truncation via gradient-based adjustment in latent/action space — so what: restores short-horizon closed-loop behavior without retraining the VLA backbone.
  • Event-triggered adaptive action horizon: the detect-and-correct loop yields long open-loop execution when dynamics are stable and short-horizon replanning when drift occurs — so what: mitigates the static-horizon trade-off between robustness and policy-call frequency.
  • Integration-first design: works as an inference-time wrapper around existing VLA models, keeping backbone weights unchanged — so what: lowers adoption cost for deployed systems that need improved robustness.
Who it's for, and trade-offs

Great fit if you run action-chunked vision-language-action policies in long-horizon or contact-rich manipulation scenarios and want to reduce failures without the expense of full closed-loop control or retraining. Look elsewhere if you require guaranteed real-time, high-frequency control (the wrapper adds inference overhead) or if you prefer end-to-end learned closed-loop controllers tuned via additional training data. VLA-Corrector improves practical robustness at inference time but does not replace benefits obtainable from retraining or higher-rate control loops.

How it works (brief)

The base VLA policy predicts an action chunk and an implicit visual evolution. LVM encodes actual camera observations into the same latent space, measures deviation from predicted evolution, and accumulates this signal to detect persistent drift. Upon a truncation event, remaining planned actions are discarded and OGG performs a short, gradient-guided replanning pass to produce corrective actions. The system thereby adapts the effective action horizon at runtime, interrupting compounding errors while retaining most efficiency gains of chunked execution.

Overall, VLA-Corrector offers a practical, inference-only path to make action-chunked VLA policies more robust in real-world manipulation tasks where small local perturbations can otherwise lead to rapid failure.

Information

  • Websitearxiv.org
  • AuthorsYi Pan, Miao Pan, Qi Lu, Jiaming Huang, Man Zhang, Siteng Huang, Xin Li, Jie Zhang, Yongliang Shen, Xuhong Zhang …
  • Published date2026/07/02

More Items

Trains AI reviewers to identify substantive, evidence-based concerns about papers. Its reinforcement-learning objective balances the precision of critiques with their coverage of known high-quality concerns.

Generates coordinated portfolio weights directly from multi-asset market histories, then refines them with portfolio-level reinforcement learning. In cryptocurrency futures tests, refinement improved net annualized return while reducing maximum drawdown.

Hugging Face

Provides first-person video of skilled manual work with synchronized motion-sensor data for robot learning and dexterous manipulation research. The first release spans 10,124 hours from 12,227 contributors, with task-level annotations.