Encore: Few-Shot Agentic Discovery of Manipulation Strategies
ENCORE turns a few robot demonstrations into frozen manipulation policies without leaking the success signal.
ENCORE gives a coding agent a few robot demonstrations distilled into multi-view keyframes, gripper events, frame strips, and full trajectories, rather than using them as training data. The agent writes a policy against a fixed perception and action API, refines it over development rollouts, and freezes it before a sealed evaluation that hides the success signal. On LIBERO-PRO, frozen programs reach 96.3% versus 89.3% for the strongest prior agentic system using the same language model; without demonstrations, no program succeeds on RoboDojo tasks with unstated goals. It also learns cube handover and cup inversion on a real bimanual robot from five demonstrations each.
- Demos are packaged as keyframes, gripper events, strips, and trajectories.
- The coding agent writes and freezes a policy before sealed evaluation.
- LIBERO-PRO frozen programs score 96.3 percent versus 89.3 percent prior.
- A real bimanual robot learns handover and cup inversion from five demos.
From the original article
1 min read at huggingface.co
Coding agents can now write, run, and debug programs with little human help. Robot tasks, however, are usually specified by a sentence that leaves out how to grasp, in what order to make contact, and what the result should look like, and an agent given only the sentence must find these details by trial and error.
By Yifan Kang, Zihan Wang, Zhiwen Fan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.