control policy
A control policy is the rule a robot follows to decide what to do at every moment. Feed it what the robot currently knows about itself and the world — its joint angles, what its camera sees, how fast it is moving — and the policy hands back an action: turn this motor, take this step, close this gripper. You can picture it as the robot's instinct or its 'if I see this, I do that' habit. The whole point of robot learning is to grow a good policy from data and experience instead of having a person spell out every if-then rule by hand.
Policies come in two flavors. A deterministic policy is a strict recipe: the same situation always produces exactly the same action, like a thermostat that switches the heater on at precisely the same temperature every time. A stochastic policy instead gives a spread of possibilities — 'mostly go left, sometimes go straight' — and then rolls dice to pick one. That little bit of built-in randomness is surprisingly useful: it lets the robot try new things while still learning, and it keeps the robot from getting predictably stuck when the strict recipe would jam it against the same wall over and over.
A walking robot's policy reads its tilt and foot positions and outputs the next push for each leg motor, twenty times a second, to keep it upright.
A policy maps what the robot senses now to what it does next.
In robot learning the policy is the thing being learned; reinforcement learning is one common way to learn it.