Reinforcement learning
Physical AI
Learning a policy to maximize expected cumulative reward through interaction data.
The reward defines the optimization objective. A high reward does not automatically mean the intended task was completed correctly. Review should check reward labels, terminal conditions and the evidence behind task outcomes. Training can use newly collected interaction or previously recorded data, depending on the method.
In practice
A robot receives a success reward only when the insertion remains seated after release, rather than when contact first occurs.