Researchers led by Qifeng Xie have built FIRE-Drive, a planner for end-to-end self-driving models that generate paths with diffusion, according to a paper on arXiv. Such models depend heavily on where their candidate paths start and on choosing between paths that look alike but differ in safety and progress.

FIRE-Drive starts from feasible candidate paths, scores each one with risk evidence from a bird's-eye view of the scene and trains on a small set of high-quality candidates. The diffusion decoder then refines plausible paths, while the scorer picks between them.

Using a ResNet34 backbone, it scored 91.5 on the NAVSIM v1 benchmark's Predictive Driver Model Score. Over 220 Bench2Drive routes driven in closed loop, it completed 42.73% of them and its driving score was 71.43.

Source: arXiv Robotics (cs.RO)

Background

Qifeng Xie

Robotics researcher

Robotics researcher.