Researchers improve robot task success with navigable video demonstrations

Researchers report a training-free method that improves robot task success by letting an AI agent consult selected parts of a demonstration video. In their preprint, Recursive Video In-Context Learning raises success from 92.6% to 96.5% on LIBERO-PRO and from 86.7% to 95.8% on LIBERO-Plus, using one demonstration per task. The method organizes grasps, releases and other sub-events into increasingly detailed levels, from task keyframes to short clips. The agent reads broad views before planning, then retrieves details for its current step during execution. The reported gains come from a system built on RPent and evaluated on those two benchmarks.