Four Studies Push End-to-End Autonomous Driving Past Sparse Scenes

End-to-end autonomous driving models often compress the world into sparse elements — objects, lane lines, and the like. Two new papers argue this representation is dangerous in crowded, occluded scenes, where missing evidence leads to poor candidate generation and risky behavior. Both propose a shift toward risk-aware occupancy, which models continuous spatial risk rather than discrete entities, to give planners a more complete picture of uncertainty.

A separate paper takes a step back, examining how these planners are evaluated. The authors unify CARLA-based leaderboards into a single benchmark and find that inconsistent metrics and routes across existing evaluations make fair comparison difficult. This work doesn't propose a new driving algorithm but a standardized testing framework, aiming to make progress in the field measurable.

A fourth paper targets the output side: waypoint-based imitation learning can produce jittery or unstable trajectories, which downstream controllers struggle to follow. The authors introduce SC-IMM-based teacher signals to stabilize trajectory predictions, improving the consistency of vehicle control commands.

Together, the papers agree that current end-to-end systems are fragile — either due to sparse perception, unstable outputs, or unclear evaluation. They differ in where they apply the fix: perception representation (risk-aware occupancy), evaluation methodology (unified CARLA benchmark), and control stability (SC-IMM teacher signals). No single paper addresses all three, but collectively they point toward a more robust pipeline.