What is a robot policy?

To understand a policy, describe its interface. What does it observe: images, joint positions, force readings, history or a language instruction? What does it output: joint targets, end-effector motion or another action representation? The same word can refer to quite different systems. The LeRobot documentation provides concrete examples of policies and their surrounding training and robot workflows.

A policy result needs a task, hardware configuration and evaluation protocol. Changing the gripper, camera location, objects or environment may change the result. Training support for a robot platform does not itself prove reliable task execution. Distinguish a policy checkpoint from the full system used in a demonstration. Inspect the inference setup, lower-level control and recovery behavior before comparing two releases.

Which version was tested? Was the test task represented in training? Was a person selecting starts, resetting objects or correcting failures? What counted as success? Were all attempts reported? A policy can have a strong task result while still depending on carefully prepared conditions. Our evaluation checklist turns these questions into a reusable record.