CoRe gives robots self-recovery with 85-point success boost

Robotic manipulation powered by vision-language-action (VLA) models is increasingly flexible, but it breaks down when something unexpected happens—a changed goal, a shifted object, or a bumped robot arm. Existing recovery methods rely on failure-specific training data, policy fine-tuning, or external corrective agents, which adds cost and risk. In a new arXiv paper, Yanyan Zhang and colleagues from multiple institutions propose CoRe (Counterfactual Realignment), a training-free framework that lets a frozen VLA model recover itself purely at inference time.

CoRe works by detecting a deviation, then imagining how the policy would have continued from a recent viable state—using synthesized observations instead of physical execution. It then minimally realigns the robot and scene to rejoin that imagined path before handing control back to the policy. This avoids physical trial-and-error, preserves task progress, and unifies recovery from both instruction changes and physical perturbations. In experiments across multiple simulators, VLA backbones, and real-world setups, CoRe improved success rates by up to 85.0 percentage points to near-nominal levels while reducing physical restorations by 42.2%, without any policy fine-tuning or failure-specific training.