DiMuRe: Diagnosis-Guided Multilevel Recovery for Robotic Task Execution
| Source: arXiv AI
Tags: DiMuRe, LIBERO-Disturb, Vision-Language-Action, Robot Recovery, Multimodal LLM
DiMuRe lets robot agents diagnose mid-task disturbances and pick a recovery level, raising complete-task success from 58.7% to 77.7% on a new disturbance benchmark, and to 81.9% with replanning.
Details
When a disturbance invalidates a robot's next action, much of its task program may still be usable. DiMuRe uses phase-specific evidence and method preconditions to choose between reusing the program and predefined recovery skills, including bounded vision-language-action grasping. After the original program exits, a multimodal LLM can generate corrective code from images and execution feedback, and local verification checks each recovery attempt. The authors also release LIBERO-Disturb, a benchmark for mid-execution disturbances supporting program-based agents and VLA policies. On 801 disturbed cases, local DiMuRe reached 77.65% success versus 58.68% for the original program, using fewer physical steps. Full DiMuRe reached 81.90% with replanning and extra step allowance.