Agent as Policy for Robotic Manipulation
A general-purpose agent ran real robot manipulation tasks without task- or environment-specific training.
The paper introduces Agent as Policy, or AGP, putting planning and execution under the agent’s control through visual interpretation, executable code, motion commands, and runtime correction. The authors tested it on real-world tasks including assembly from human videos, block construction from goal images, dice flipping, targeted throwing, and bimanual towel folding. In assembly, block construction, and dice flipping, it succeeded in at least eight of ten trials for each evaluated configuration. Reusing saved procedures and programs reduced execution time in repeated trials. HF Daily Papers' note
The paper introduces Agent as Policy, or AGP, putting planning and execution under the agent’s control through visual interpretation, executable code, motion commands, and runtime correction. The authors tested it on real-world tasks including assembly from human videos, block construction from goal images, dice flipping, targeted throwing, and bimanual towel folding. In assembly, block construction, and dice flipping, it succeeded in at least eight of ten trials for each evaluated configuration. Reusing saved procedures and programs reduced execution time in repeated trials. HF Daily Papers' note
score 5