VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon
VLA-Corrector addresses latency issues in VLA models by creating a lightweight corrective inference framework for action-chunked policies. It uses a Latent-space Vision Monitor to detect visual deviations, triggering truncation and Online Gradient Guidance to steer the robot back on track. This transforms fixed action horizons into adaptive ones, improving efficiency and robustness without retraining the main policy.