Encore: Few-Shot Agentic Discovery of Manipulation Strategies
ENCORE has coding agents infer robot manipulation details from a handful of demonstrations, then lock the resulting policy before evaluation.
The method turns each demo into evidence packs with keyframes, gripper events, frame strips, and trajectories, rather than using them as training data. A coding agent reads those packs, writes a policy against fixed perception and action APIs, and refines it over a few development rollouts. On LIBERO-PRO, frozen programs reached 96.3% versus 89.3% for the strongest prior agentic system using the same language model. The paper also reports real bimanual robot runs for cube handover and cup inversion from five demonstrations each. HF Daily Papers' note
The method turns each demo into evidence packs with keyframes, gripper events, frame strips, and trajectories, rather than using them as training data. A coding agent reads those packs, writes a policy against fixed perception and action APIs, and refines it over a few development rollouts. On LIBERO-PRO, frozen programs reached 96.3% versus 89.3% for the strongest prior agentic system using the same language model. The paper also reports real bimanual robot runs for cube handover and cup inversion from five demonstrations each. HF Daily Papers' note
score 5