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

VLA-Corrector is a lightweight corrective inference framework for action-chunked vision-language-action policies. It adds a Latent-space Vision Monitor (LVM) to compare predicted and observed visual feature evolution during execution. When persistent deviation is detected, the system truncates the remaining stale actions and invokes corrective replanning through Online Gradient Guidance (OGG). This creates an event-triggered adaptive action horizon: reliable chunks can execute for longer, while drifting executions trigger shorter-horizon corrections. The authors state that the method requires no modification or retraining of the VLA backbone.