Can VLMs Diagnose and Recover from VLA Manipulation Faults?

Existing VLA models frequently fail in robotic manipulation tasks, with poorly structured fault types that often require expert diagnosis. While VLMs offer strong explanatory capabilities, their effectiveness in assisting VLAs is limited by their unclear role in diagnostics and inadequate collaboration mechanisms. To address this, we introduce VLA-FixBench, a fault evaluation dataset that spans perception, planning, and control failures, and provides annotations for task stages, fault types, and spatiotemporal repair strategies. We further propose FaultEval, a static-to-dynamic-to-real evaluation framework that benchmarks 20 VLMs across multiple fault-related dimensions. Building on these insights, we design a VLM–VLA collaboration mechanism that localizes spatiotemporal deviations and rolls back task execution to enable targeted recovery. Experiments show that FaultEval reliably characterizes VLM-based closed-loop diagnosis and repair. The upper-bound analysis using human expert intervention shows that an idealized feedback loop can improve task success rates by 13\% on LIBERO and 35\% on real-world robots. Our code, benchmark, and project page will be publicly released at: https://kakigo.github.io/VLA-FixBench/

Robots powered by modern AI can perform many manipulation tasks, such as picking up objects, opening containers, or placing items in the right location. However, these robots still often fail during execution, and understanding why they fail usually requires human experts. This paper studies whether vision-language AI systems can help robots recognize failures, explain what went wrong, and suggest how to recover. We introduce VLA-FixBench, a dataset of robot manipulation failures covering problems in perception, planning, and control. We also build an evaluation framework to test how well different vision-language models can detect failures, locate when and where they happen, and provide useful recovery suggestions. Based on these findings, we design a robot recovery mechanism that can stop execution, roll back to an earlier safe step, and apply a corrective action. Our results show that current AI models are still limited in reliable robot recovery, but accurate human-level feedback can substantially improve task success. This suggests that better failure diagnosis and recovery could make future robotic systems safer and more reliable.