Summary

This paper introduces CoRe (Counterfactual Realignment), a training-free, inference-time framework that recovers a frozen VLA policy from online disruptions — goal changes, scene shifts, robot-state errors — without needing failure data or retraining. Upon detecting a deviation, CoRe imagines how the policy would have continued from a recent viable state using synthesized observations, then minimally realigns the robot/scene to rejoin that imagined trajectory.

Key Contributions

  • A training-free recovery mechanism that requires no failure-case dataset and no fine-tuning of the base policy, in contrast to most VLA recovery methods which retrain on curated failure/recovery pairs.
  • A counterfactual framing: instead of directly correcting the current bad state, the method reconstructs what the policy “would have done” from a still-viable earlier state and steers execution back toward that imagined continuation.
  • Targets a class of disruptions (goal changes, scene shifts, state errors) broader than the contact-slip or grasp-failure recovery scope common in prior work.

Strengths

  • Training-free deployment is a genuine practical advantage — it can be layered onto any existing frozen VLA without a separate fine-tuning pass or curated failure corpus.
  • Addressing goal/scene-level disruptions, not just low-level execution slips, covers a failure mode (environment or instruction changing mid-task) that most recovery papers in this vault do not directly target.

Weaknesses

  • The method depends on synthesizing plausible “imagined” observations for a counterfactual trajectory; the fidelity of this synthesis under significant scene change is the crux of the approach and is not detailed in available coverage.
  • Detecting when a deviation warrants counterfactual realignment (versus normal task variation) is itself a hard problem — false positives could trigger unnecessary and disruptive realignment.

Open Questions

  • How does CoRe distinguish a genuine disruption from ordinary stochastic task variation, and what is the false-positive rate in practice?
  • Does the “minimal realignment” step have failure modes where scene edits themselves introduce new physical inconsistencies the base policy then mishandles?

Significance

Adds a training-free entry to the fast-growing body of test-time recovery and adaptation work for VLA policies (alongside FAR, HAVE, and Retrieve-then-Steer already logged here), notable specifically for avoiding any dependency on curated failure data.