VLA-Corrector logo VLA-Corrector
EN 中文

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

A lightweight inference-time corrector for mitigating open-loop blind spots in action-chunked VLA policies.

Yi Pan1, Miao Pan1, Qi Lu1, Jiaming Huang1, Man Zhang1, Siteng Huang2, Xin Li2, Jie Zhang1, Yongliang Shen1, Xuhong Zhang1, Wenqi Zhang1

1Zhejiang University   ·   2Alibaba DAMO Academy

panyi0304@gmail.com zhangwenqi@zju.edu.cn

Code README arXiv 2607.01804 BibTeX
Presentation
Open full
Open-loop versus closed-loop execution comparison
Open-loop versus closed-loop execution. Under a long action horizon, a robot may continue stale actions after a deviation; strict per-step replanning improves reactivity but is expensive for VLA policies.

Real-Robot Demonstrations

Silent real-robot clips on AgileX PiPER tasks. Each clip shows a human perturbation during execution and the policy recovering through the correction pathway.

Drawer alignment Human moves the drawer while the robot aligns to the top corner.
Block to blue bowl Human moves the target bowl during execution.
Block to white bowl Human perturbs the object before successful placement.

Abstract

VLA-Corrector turns fixed action chunks into an adaptive execution process: keep long-horizon efficiency when execution is stable, and interrupt stale actions when visual dynamics drift.

Problem

Chunked VLA policies can keep executing queued actions after perturbations.

Detector

A Latent-space Vision Monitor compares predicted and observed visual feature evolution.

Recovery

Persistent mismatch triggers truncation and OGG-guided corrective replanning.

Motivation

Action chunking reduces expensive VLA calls, but it also creates a blind spot between replanning events. VLA-Corrector targets that gap without retraining the full VLA.

01

Monitor latent dynamics

Predict short-horizon visual residuals and compare them with incoming observations.

02

Interrupt stale chunks

Stop executing the queued chunk when mismatch persists across consecutive steps.

03

Replan only when needed

Use OGG-guided recovery queries instead of forcing per-step VLA replanning.

Method Overview

VLA-Corrector keeps the action-chunked VLA interface intact and adds a lightweight monitoring-and-recovery path at inference time.

VLA-Corrector method overview
Overview of VLA-Corrector. Starting from a standard chunked VLA pipeline, the framework adds LVM detection, interrupt-triggered truncation, and OGG-guided corrective replanning.

The external corrector is trained after a VLA policy has been obtained. The VLA backbone is frozen, and its visual encoder extracts latent representations from demonstration trajectories. The corrector predicts short-horizon latent residuals induced by executed actions.

During deployment, LVM compares predicted and observed latent residuals. Persistent inconsistency triggers an interrupt event, discards the stale action queue, and applies OGG to the next recovery query.

Experiments

The paper evaluates VLA-Corrector on MetaWorld, LIBERO, and a real AgileX PiPER 6-DoF robot. PI0.5 is used as the main backbone, with SmolVLA and X-VLA included for cross-architecture evaluation.

  • Benchmarks: MetaWorld, LIBERO, and AgileX PiPER real-world tasks.
  • Backbones: PI0.5, SmolVLA, and X-VLA.
  • Metrics: task success, policy calls, recovery rate, and inference overhead.
Performance efficiency trade-off across fixed action horizons
Performance-efficiency trade-off. Smaller horizons improve responsiveness, while larger horizons preserve chunking efficiency but widen the open-loop blind spot.

Main Results

Setting Baseline + VLA-Corrector Reported change
MetaWorld, PI0.5 avg. success 48.70% 64.35% +15.65 points
MetaWorld, SmolVLA avg. success 61.90% 66.65% +4.75 points
MetaWorld, X-VLA avg. success 55.55% 59.60% +4.05 points
LIBERO, PI0.5 few-shot avg. success 94.00% 97.80% +3.80 points
AgileX PiPER real-world avg. success 55.6% 73.3% +17.7 points

In the reported MetaWorld component ablation, truncation alone improves average success from 48.70% to 60.35%, and truncation plus OGG reaches 64.35%. The paper also reports that 83.7% of truncations occur in manually labeled critical phases.

Analysis

The paper further studies when interrupts occur and how the correction pathway recovers from stale chunks.

Task-phase analysis of LVM-triggered truncation
LVM-triggered truncations concentrate in critical phases such as precise grasping and alignment.
Controlled recovery case
In a controlled recovery case, VLA-Corrector truncates stale actions, replans with OGG, and completes the task where the monitored baseline continues the original chunk and fails.

Code

The code repository contains the LeRobot-based implementation, modified VLA evaluation entry points, and the latent dynamics corrector training modules. Datasets, raw demonstration data, pretrained weights, fine-tuned checkpoints, trained corrector checkpoints, outputs, logs, and caches are not included. The project page only includes compressed silent real-robot clips for visualization.

See the GitHub README for installation, training, and evaluation commands.

Citation

The paper is available on arXiv: 2607.01804.

@article{pan2026vla,
  title={VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon},
  author={Pan, Yi and Pan, Miao and Lu, Qi and Huang, Jiaming and Zhang, Man and Huang, Siteng and Li, Xin and Zhang, Jie and Shen, Yongliang and Zhang, Xuhong and others},
  journal={arXiv preprint arXiv:2607.01804},
  year={2026}
}