Planning-Oriented End-to-End Autonomous Driving: Architectures, Evaluation, and Emerging Paradigms
A recent survey examines the shift in end-to-end autonomous driving research from direct camera-to-control regression models toward planning-oriented architectures. These systems incorporate structured scene representations, generate trajectory-level outputs, and adopt more realistic evaluation protocols. The review traces developments across behavior cloning, conditional imitation learning, privileged distillation, bird's-eye-view and vectorized planning, and unified perception-prediction-planning pipelines.
This transition is relevant for robotics and automation coordination because it emphasizes integrated modules that connect perception directly to motion planning. Such approaches appear in systems that must operate alongside other agents, whether in vehicle fleets, drone swarms, or warehouse automation, where consistent trajectory outputs can support synchronization across multiple platforms.
The survey notes that evaluation protocols in the field have moved toward closed-loop and real-world metrics, though direct comparisons across methods remain limited by differing simulation environments and sensor setups.