FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models

2 min read

Manipulation datasets mostly hold successful trajectories, and the few failure datasets explain failures in text that a VLA cannot act on. FailSafe generates failures automatically in simulators that have motion planning, and pairs each one with an executable recovery action. The resulting data fine-tunes LLaVA-OneVision-7B into FailSafe-VLM, which works beside a VLA: every 10 steps it takes control, checks for a failure, and if it finds one outputs an action the robot can execute directly, then hands control back. First posted in October 2025, presented at IROS 2026 in the session on securing VLA robots against failure.

Findings

  • Improves three VLAs (π0-FAST, OpenVLA and OpenVLA-OFT) by up to 22.6% on average across several ManiSkill tasks.
  • Generalises across spatial configurations, camera viewpoints, objects and robot embodiments, per the abstract.

Limits

  • Recovery is at the level of motion. Object-level errors, such as picking the wrong object, are not corrected (authors).
  • The pipeline needs a simulator with motion planning to make failures and their fixes, so failure types are those a planner can produce and undo.
  • A second 7B model runs beside the policy, the cost Progress Probes and Backtrack try to avoid.

Relevance

FailSafe is the “external judge that also proposes the fix” design in failure-recovery. Compared with Backtrack, it learns what to do after a failure rather than only rewinding. It belongs in comparison-failure-detectors as a VLM-based detector with recovery attached, and its judge would be exposed to the coverage shift problem as the policy it watches improves.