Select Website for a short introduction to each paper.
-
Can VLMs Diagnose and Recover from VLA Manipulation Faults?
Bowen Yan, Jiahao Xiao, Kehui Liu, and 8 more authors
In International Conference on Machine Learning (ICML), 2026
VLA-FixBench contains 6,034 task-execution episodes for studying VLM-assisted VLA fault diagnosis and recovery. FaultEval evaluates 20 VLMs from static diagnosis to simulation and real-robot recovery. A detect-rollback-correct interface exposes the gap between diagnostic reasoning and reliable intervention; real-robot success reaches 65 percent with human-expert intervention, compared with 30 percent for the VLA-only baseline and at most 25 percent for the evaluated automatic VLM judgers.