Driving planner preprint reports energy use below 2% of neural-network baselines

Researchers report that SDPAD, an autonomous driving planner using sparse integer signals, consumed 69.9 mJ on the nuScenes benchmark-less than 2% of recent conventional neural-network baselines. The preprint reports an average trajectory error of 0.40 m and a collision rate of 0.12%, comparable to strong conventional planners. The system converts a pretrained perception network into integer-spike form, transforms camera views into an overhead scene representation and produces a planned trajectory in one forward pass. In closed-loop evaluation on NAVSIM's navtest split, it scored 86.3 PDMS, exceeding the earlier spiking planner SAD by 4.3 points. These results concern benchmark evaluations.