inverse RL vs. imitation

These are the two answers to one question: given expert behavior, what should you learn? Imitation learning (including behavioral cloning) learns the behavior directly — a policy mapping states to actions, copying the what. Inverse RL learns the explanation — a reward function capturing the why — and then plans or runs RL against it to act. One reaches for the action; the other reaches for the goal behind it.

The trade-off is portability versus directness. A cloned policy is simple and needs no environment model, but it is locked to the world it was demonstrated in and cannot exceed the expert. A recovered reward is harder to obtain and ambiguous, but it transfers: change the dynamics, the robot body, or the start state, and you can re-optimize the same goal, potentially surpassing the demonstrator on it.

In practice the line blurs. Adversarial methods like GAIL imitate distributions without naming a reward, AIRL extracts a transferable reward from the same game, and demonstration-guided RL bolts imitation onto reward optimization — so the honest framing is a spectrum from copying actions to recovering intent.