Imitation Learning vs RL Distribution Shift Debate

ftm_guney · x · 2026-07-15

The author believes that everyone advocates adding RL on top of imitation learning to solve distribution shift; but in reality, imitation learning agents like TFv5/6 and DrivoR still dominate in open benchmarks. He thus asks: is the so-called 'distribution shift' problem that RL claims to solve really as severe as people say?

Related event: Why imitation learning still tops RL on benchmarks(2 posts)→

Original post →

More from Research

Research channel →