This research explores how well AI agents, trained by watching expert demonstrations, can adapt to unexpected changes in their environment. I conducted a comprehensive study comparing three key imitation learning algorithms (BC, GAIL, and AIRL) across several simulated robotic tasks with altered physics and goals. The findings reveal the limitations of current methods in generalizing to novel situations and highlight the challenges of transferring a learned understanding of a task's objective.

This research investigates the "sim-to-real" gap, where AI agents trained in simulation falter in the real world. I benchmarked model-based (SHAC) versus model-free (SAC, PPO) reinforcement learning algorithms on their ability to adapt to changes in physics and sensor noise in the classic Pendulum control task. My experiments revealed that the model-free SAC algorithm was not only more robust to these variations but also learned an expert policy five times faster than the other methods.