A driving planner organizes candidate paths by motion
In plain words
ManeuverFlow is a method for planning an autonomous car’s path. It makes several possible paths, groups them by the kind of motion they represent, checks each path against one prediction of the scene ahead, and adjusts some paths before choosing one.
In tests against logged driving scenes, its LiDAR-assisted version received a mean score of 90.58 on NAVSIM v1’s composite driving measure. The result applies to that test setting; it does not show how the planner performs in real traffic.
The research
When a driving scene allows several plausible paths, how can a planner cover distinct motions with a limited candidate set and still judge what may happen along each route?
ManeuverFlow addresses this by organizing path generation around motion categories and checking candidates against a shared prediction of the scene ahead.
ManeuverFlow groups candidate driving paths by motion, scores them using evidence from current and predicted scenes, refines selected paths, then chooses one.
ManeuverFlow is an end-to-end driving planner that generates paths for distinct trajectory-based motion categories, then checks and refines selected paths using evidence from current and predicted scenes.
NAVSIM evaluates planned paths against logged future scenes. On its v1 test set, the LiDAR-assisted version received a mean composite score of 90.58; on v2, its v1-trained selector received a mean extended score of 90.14 without v2-specific tuning.
These results come from NAVSIM’s logged-scene evaluation and WorldEngine scenarios reconstructed from NAVSIM logs.
They do not establish performance in independent driving domains or real vehicles, and the shared future-scene prediction does not model how other drivers would respond to alternative ego plans.
Analysis
Deep
- Authors
- Yuhao Wei, Yong Chen, Chen Yang, Yingchen Wang, Delin Ouyang
- License
- arXiv non-exclusive