A learning model (e.g. a neural network or RL policy) is trained on simulated or real data to map state observations to control actions; the policy is optimized against a reward/objective function and refined iteratively, often with simulation-to-reality transfer (sim-to-real).
For many systems (e.g. dynamic locomotion, contact, dexterous manipulation) an accurate analytical model is hard to derive; learning-based control obtains effective control strategies directly from data.
