LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving

ArXi:2512.20563v2 Announce Type: replace-cross Simulators can generate virtually unlimited driving data, yet imitation learning policies in simulation still struggle to achieve robust closed-loop performance. Motivated by this gap, we empirically study how misalignment between privileged expert nstrations and sensor-based student observations can limit the effectiveness of imitation learning. precisely, experts have significantly higher visibility (e.g., ignoring occlusions) and far lower uncertainty (e.g., knowing other vehicles' actions), making them difficult to imitate reliably.