Imitation Learning

Imitation learning bypasses complex reward engineering by training autonomous agents to mimic expert demonstrations directly.

Instead of relying on trial-and-error reward functions (which are notoriously difficult to design for complex environments), imitation learning trains machine learning models directly on expert demonstration data. By mapping states to actions using techniques like behavior cloning or inverse reinforcement learning, agents can rapidly acquire complex skills. This approach is highly effective in robotics and autonomous driving, where physical exploration is costly and human demonstrations (such as steering angles or robotic arm trajectories) provide a direct, safe path to policy optimization.

https://arxiv.org/abs/1801.06503