A different way to get a control law
Every controller you've studied so far on this site was written by a human. A PID controller is a formula an engineer derives from feedback theory and then tunes. A motion planner is an algorithm an engineer designs to search a space of paths. In both cases, a person looks at the physics of the problem, reasons about it, and writes down the logic that turns sensor readings into motor commands. The robot is executing a rule someone else figured out.
Reinforcement learning flips this. Instead of an engineer writing the control law, an RL agent learns a policy, a mapping from observed state to action, by repeatedly acting in an environment and receiving a reward signal that scores how good or bad each outcome was. Over many, many attempts, sometimes millions, the agent adjusts its policy to increasingly favor actions that led to higher reward in the past. Nobody writes down 'if the arm is tilted this way, apply this torque.' The robot discovers that relationship itself, purely from the consequences of its own actions.