Autonomous driving planners typically generate a set of candidate trajectories, score each one, and select the highest-scoring option. The scorer, however, is applied only after the candidates are produced, so it cannot fix a weak candidate set. A new arXiv paper, "Test-Time Trajectory Optimization for Autonomous Driving," proposes treating the scorer as a learned reward function and searching for trajectories that maximize it.
The method, called TOAD, runs the Cross-Entropy Method at test time, warm-started from the planner's own proposals. Because it requires no retraining, it can be added to existing planners as a plug-and-play module. The authors report improvements across seven base planners on the NAVSIM-v1 and NAVSIM-v2 benchmarks, reaching 94.7 PDMS and 56.3 EPDMS respectively, as well as on the closed-loop HUGSIM benchmark.
The approach addresses a common bottleneck in end-to-end planning: the quality of the final trajectory is bounded by the diversity and coverage of the initial candidate set. By optimizing at test time, TOAD shifts some of the burden from proposal generation to search. The code is publicly available on the project page.