Can Spiking Neural Networks Finally Match Dense Networks in Autonomous Driving Planning?

The Energy Gap That Kept Spikes Out of Planning

End-to-end autonomous driving demands trajectory planners that are both highly accurate and cheap enough for edge deployment. Spiking neural networks promise orders-of-magnitude energy savings but still lag far behind in planning accuracy. This gap has kept spikes out of the planning stack until now.

How SDPAD Converts a Trained Network Into Integer Spikes

SDPAD converts a pre-trained ANN perception stack into integer-spike form via quantized ANN2SNN conversion. The pipeline lifts multi-view images into bird's-eye-view space with a spike-driven-max depth distribution called Spike-3D-Lift. A spiking query transformer named Spike-QFormer then plans the trajectory. Ego queries and agent queries distilled from the BEV scene are fused by learnable waypoint queries via cross-attention. Map queries join the fusion in a second step. Deformable spike-cross-attention refinement follows. Every operation is gated by integer spikes. Inference is a single feed-forward pass without temporal simulation loops. No recurrent time steps are required. State-of-the-art ANN planners meet the accuracy requirement at the cost of heavy dense computation.

Benchmarks

On the nuScenes open-loop benchmark SDPAD achieves an average L2 error of 0.40 m and a collision rate of 0.12%. Energy consumption measures 69.9 mJ, less than 2% of recent ANN baselines. In closed-loop evaluation on the NAVSIM navtest split SDPAD reaches 86.3 PDMS, beating the previous SNN planner SAD by 4.3 points while matching mainstream ANN planners at a fraction of their energy.

What Remains Unknown

The paper does not report results on real vehicle hardware. It does not describe how the integer-spike conversion affects latency under thermal throttling. Whether the approach scales to higher sensor counts or longer planning horizons remains open. The quantized ANN2SNN conversion step may introduce quantization error not fully characterized in the benchmarks. Real-world deployment will require validation beyond simulation.