ENCORE: Coding Agents Write Robot Policies From 5 Demos, Hitting 96.3% on LIBERO-PRO
2077AIDataFoundation · hf · 2026-10-12
ENCORE, from MIT, 2077AI, and collaborators, treats a few demonstrations as evidence to read rather than training data, letting a coding agent write robot manipulation policy programs.
- A deterministic builder distills each demo into a pack of multi-view keyframes, gripper events, frame strips, and full trajectories; the agent writes a policy against a fixed perception/action API, refines it over a few dev rollouts, then freezes it before sealed evaluation.
- On LIBERO-PRO, the first program already succeeds on half of perturbed tasks with demos, and frozen programs beat the strongest prior agentic system, 96.3% vs 89.3%. On RoboDojo tasks with unstated goals, no program succeeds without demos.
- ENCORE also runs on a real bimanual robot, learning cube handover and cup inversion from five demos each.
More from Embodied
- Dev builds Claude-coded 'daily wearable' firmware packing 27 games and 16 utilities — ezshine · 2026-10-12
- China's driverless buses now ship without steering wheels — CurieuxExplorer · 2026-10-12
- HomeBody gives VLMs persistent spatial memory: Unitree G1 cleans a kitchen with no task-specific training — rsasaki0109 · 2026-10-12
- ARTISAN: Multi-Agent Pipeline Generates Physically Grounded Articulated 3D Assets with New Benchmark — elliottszwu · 2026-10-12
- $84,000 Boston Dynamics robot washes a lioness, sparks animal-intervention debate — CurieuxExplorer · 2026-10-12
- San Francisco robot fight club levels up: 200-pound humanoids duel with steel blades — cixliv · 2026-10-12