Part 3/15:
A key insight from Dr. Fan's research emphasizes that optimizing AI policies solely in simulation often leads to poor real-world performance. The core challenge lies in the differences—or discrepancies—between simulation and reality, which can hinder an AI agent's effectiveness once deployed physically. To bridge this gap, engineers collect small amounts of real-world data to enhance the fidelity of their simulators. This process iteratively improves the accuracy of the virtual environment, leading to more robust AI deployment in actual hardware.