End to End Autonomous Driving via Occupancy and Motion Flow

Many existing end-to-end autonomous driving methods involve reinforcement learning or multi-stage discrete task pipelines. Reinforcement learning approaches lack deterministic interpretability, and multi-stage discrete task pipelines rely on dense scene representations (detection, tracking, segmentation) and suffer from significant computational redundancy. Additionally, previous neural motion planners often treated perception and planning as independent components, leading to compromised trajectory accuracy and threatening driving safety. In this paper, we propose OFAD(Occupancy and motion Flow for Autonomous Driving), a novel end-to-end learning paradigm based on semantic occupancy and motion flow for perception, prediction and motion planning in autonomous vehicles, while generating interpretable intermediate representations. Furthermore, motion flow prediction is explicitly used as a cost function in the motion planning process, ensuring consistency between perception and planning. Experimental results on the large-scale manually-driven dataset, nuscenes, demonstrate that OFAD significantly outperforms previous planners in mimicking human driving, generates safer trajectories, and provides meaningful results for drivers.