Autonomous planners no longer keep one trajectory: eight maneuver hypotheses in one step, 90.58 PDMS on NAVSIM

The planner now keeps candidate trajectories grouped by interpretable motion maneuver instead of letting a set of queries cluster freely in trajectory space. Learned queries without explicit mode-level structure may concentrate in similar regions of trajectory space, producing redundant candidates and risking mode collapse, while fixed anchors may spend part of the candidate budget on modes weakly relevant to the current scene. In the complete pipeline, the maneuver-aligned M-GMP-8 prior exceeds the KMeans-8 control by 0.54 PDMS and 0.83 EP, indicating that maneuver-aligned prior organization itself contributes.

Each candidate trajectory is scored not only from the current frame but by querying predicted future scene states along that path. Current-only scorers do not explicitly query predicted future scene states along candidate paths, and WoTE provides such evidence through one future rollout per candidate, repeating future BEV prediction across candidates. At a fixed candidate budget, adding predicted-future evidence raises PDMS by 1.58 and EP by 2.29; after ECFR, the complete model still exceeds the version without the future branch by 0.55 PDMS and 0.86 EP.

Selected candidates are locally refined in a path-aligned Frenet frame and then re-scored jointly with the unrefined candidates. Candidate evaluation and trajectory refinement were previously strengthened separately, and retained unrefined alternatives were rarely compared directly with refined trajectories under the same future-conditioned evaluator. ECFR adds 0.65 PDMS and 0.95 EP to the future-aware pipeline at a cost of 0.10 points of NC, and also improves the current-only variant.

The pipeline reaches 90.58 PDMS and 90.14 EPDMS on NAVSIM v1/v2 navtest with a ResNet-34 backbone, and its camera-only LTF variant reaches 75.19 and 74.58 PDMS∗ on all 288 publicly released high-risk WorldEngine scenarios, under logged replay and IDM-controlled traffic respectively. It suits readers following end-to-end planning candidate generation and selection, especially engineering teams that must organize multimodal hypotheses under a fixed candidate budget.

Maneuvers are defined as trajectory-space motion components partitioned by rules over displacement, mean heading, and speed trend, not as a separate map-semantic annotation. Top-three routing may still miss rare or compositional behavior, and the candidate-independent future field does not model interactive responses to alternative ego plans. NAVSIM v2 reuses v1-trained checkpoints and selector, and WorldEngine scenarios are reconstructed from navtest logs, so both test fixed-selector transfer and closed-loop replanning rather than independent-domain or real-vehicle generalization.