VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon
VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon
The paper proposes VLA-Corrector, a framework that enhances the reliability of VLA policies by introducing a Latent-space Vision Monitor (LVM) to detect visual deviations and trigger corrective replanning. This approach allows for adaptive action horizons, improving robustness in contact-rich robotic tasks without retraining the policy backbone.
This paper introduces VLA-Corrector, a lightweight framework that improves the robustness of action-chunked VLA policies by detecting and correcting visual deviations in real-time, without retraining the underlying policy.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org