Event

Failure Taxonomy and Single-Sentence Correction Method for Frozen VLA Policies

1 reports · 1 sources

Owen Kwon and co-authors annotated the failures of OpenVLA-OFT and pi0.5 one by one on out-of-distribution LIBERO-90 tasks, induced four failure modes, and claim that only one of them, "execution misalignment," can be repaired by a single retrospective natural-language correction. Rerunning the failed trajectories in simulation, the method recovered 65 of 139 in-distribution failures and 33 of 328 out-of-distribution failures. On a real xArm7, after the robot base was moved, under two same-object conditions pick-and-place improved from 10/30 to 30/30, and with a new object from 2/30 to 24/30. The authors state that two long-horizon tasks remained at 0/30 with or without correction, consistent with the expectations of their taxonomy. The authors also report that a VLM that relied only on simple prompting, watched the same failures, and wrote corrections of the same kind never recovered in up to 10 attempts.