What a "Policy" Means in Robot Learning

One of the most-used terms in modern robotics, and it just means something fairly simple.

In robot learning, a 'policy' is simply the function (usually a neural network) that maps whatever the robot currently senses — camera images, joint positions, force readings — to the action it should take next, whether that's a joint torque command, a target position, or something higher-level.

Why the term is used so broadly

It's a deliberately general word that applies whether the underlying decision-making came from reinforcement learning, imitation learning, or even a hand-coded rule system — anything that consistently turns sensor input into action output can be called a policy, which is why you'll hear it used across very different kinds of robot software.

Policy vs. planner

A policy typically reacts moment to moment based on current input, while a planner computes an explicit sequence of future actions ahead of time. Many modern systems blend the two — a planner sets a rough intended path, and a lower-level policy handles the moment-to-moment execution and disturbance rejection along it.