A new arXiv paper proposes Geometry-Aware Token Selection (GATS), which uses LiDAR and radar measurements to choose image patches before an autonomous-driving model processes them. The authors report that at a 10% token budget, GATS used 24 GFLOPs—18 times fewer than the dense baseline—while scoring better on nuScenes and NAVSIM. On 739 intent-critical scenarios, they report a 0.09% collision rate, compared with 0.34% for the dense model. These are results reported in the paper; the supplied abstract does not establish real-world driving performance.
Why it matters
For autonomous-driving researchers, the reported compute reduction suggests a way to process less camera data without losing benchmark performance, but its effect on real roads remains unknown.
