"Reinforcement learning faces major challenges outside controlled environments because repeated failures can be costly or dangerous, accurate simulations are difficult to build, and learned policies may fail to generalize to unfamiliar conditions."
In pages 93-94 of Artificial General Intelligence, Togelius explains why reinforcement learning alone is unlikely to provide a straightforward path to AGI. Learning through repeated trial and error is practical in games and fast simulations but becomes problematic in physical environments such as autonomous driving, where failures carry real costs. He also highlights overfitting and poor generalization when conditions change, including unfamiliar roads, vehicles, weather, or obstacles.
View Source