A new coding agent learns to grasp, flip, and hand off objects after watching just a few demonstrations, no trial-and-error required.
The system, called ENCORE, is described in an arXiv preprint titled "Encore: Few-Shot Agentic Discovery of Manipulation Strategies" (arXiv:2609.37359), posted September 30, 2026, and not yet peer reviewed. A builder tool breaks each demonstration into multi-view keyframes, gripper events, frame strips, and a full trajectory. A coding agent studies that pack, writes a policy program against a fixed perception-and-action API, refines it over a handful of test runs, then freezes it before a sealed evaluation. On the LIBERO-PRO benchmark, the agent's very first program already solved half of the perturbed tasks with demos, and one task with none at all; the frozen programs hit 96.3% success, ahead of the strongest prior agentic system using the same language model at 89.3%. On RoboDojo tasks, where the goal is left unstated, no program worked without demonstrations. On a real bimanual robot, it learned cube handover and cup inversion from five demonstrations each.
Most robot instructions are one vague sentence, and agents typically fill in the missing grasp order and success criteria through expensive trial and error. ENCORE instead treats a few demos as evidence to reason about, then writes inspectable code rather than training an opaque policy network. That distinction matters: auditable code is easier to debug and correct than a black-box weight matrix.
Still, this is one preprint, not yet peer reviewed, tested on two benchmarks and a single lab robot with five demonstrations apiece - promising, but nowhere near proof it handles the clutter and novelty of a real warehouse floor.