A spiking neural network can now plan a self-driving car's route about as well as conventional AI models - using a sliver of the power.
Researchers built SDPAD, a planning system that runs entirely on spikes, the brief on-off pulses that spiking neural networks use instead of the dense math typical AI models rely on. The team converted a pre-trained perception model into spike form, stitched camera views into a bird's-eye map, and fed that into a spiking transformer that outputs a driving path in a single pass, no loops required. On the nuScenes benchmark, SDPAD posted a 0.40-meter average error and a 0.12% collision rate, in the same range as strong conventional planners, while burning 69.9 millijoules - under 2% of what recent ANN-based systems use. In closed-loop testing on NAVSIM, it scored 86.3 PDMS, beating the previous best spiking planner by 4.3 points.
That efficiency gap matters because planning, not perception, is usually the computational bottleneck in self-driving stacks, and it is exactly the part that has to run on power-constrained hardware inside a vehicle. If spiking approaches can hit accuracy parity rather than lag behind as a curiosity, that reshapes the calculus for edge deployment, where every milliwatt affects battery life, heat, and hardware cost.
It is one benchmark result, not a shipped product - nuScenes and NAVSIM are standardized tests, not city traffic, and the paper does not address how this handles the harder real-world driving failures that haunt even today's conventional systems.