3D Occupancy-Guided BEV Learning Enhances Robust End-To-End Trajectory Planning

This study evaluates the contribution of voxel-level three-dimensional occupancy supervision to end-to-end trajectory planning for autonomous driving. A two-head multi-task network is developed in which semantic occupancy prediction and trajectory planning share a common bird’s-eye-view (BEV) representation. Training is conducted in two stages. First, geometry-aware BEV features are learned with a voxel-wise cross-entropy loss. Second, the planning head is then activated, and the two tasks are jointly optimized through a weighted loss, where the loss functions are designed to prioritize planning as the primary objective while allowing gradients from occupancy prediction to refine the shared features. Experiments on the Occ3D-nuScenes benchmark show that the proposed model attains 36.6 % mIoU for occupancy prediction, decreases the 3 s-horizon trajectory error by 2 %, and reduces the collision rate by 2 % relative to a planning-only baseline, without incurring additional inference latency. The results substantiate that 3-D occupancy prediction provides measurable benefits to end-to-end autonomous-driving performance.