Researchers introduce DiMuRe, a system that helps robots recover when a disruption makes their next action unsuitable. It uses evidence from the robot’s current task phase to decide whether to reuse the existing program or try a predefined recovery skill; an AI model can also generate corrective code after the program exits.
On 801 scheduled disturbed cases in the Encore development setup, the paper reports 77.65% complete-task success for Local DiMuRe, versus 58.68% for the original program under the same physical-step allowance. Full DiMuRe reached 81.90%, but with extra steps and terminal replanning. The researchers also introduce LIBERO-Disturb, a benchmark for testing recovery.
Why it matters
For robotics researchers testing task reliability, LIBERO-Disturb offers a benchmark focused on mid-task disruptions; whether the reported gains carry over to other robots or real-world settings is not established here.
