RECTOR: Rule-Based Reranking for Safer Autonomous Driving Trajectory Selection
A new post-generation reranking layer called RECTOR (Rule-Enforced Constrained Trajectory Orchestrator) prioritizes safety, legality, road norms, and comfort over model confidence when selecting a trajectory from a multi-modal candidate set in autonomous driving stacks. RECTOR scores candidates against a tiered rulebook (Safety > Legal > Road > Comfort) using differentiable proxies and a scene-conditioned applicability mechanism, then selects via a deterministic ε-lexicographic rule that preserves cross-tier priority without retraining the predictor. Tested on the Waymo Open Motion Dataset validation_interactive split (43,219 augmented instances, K=6) under Protocol B with a 28-rule proxy catalog and oracle applicability, rule-aware selection reduced Safety+Legal violations from 28.58% to 20.42% and total violations from 40.32% to 32.41% compared to confidence-only selection on the same candidates.
Key facts
- RECTOR is a post-generation reranking layer for autonomous driving trajectory selection.
- It uses a tiered rulebook: Safety > Legal > Road > Comfort.
- Scoring uses differentiable proxies and scene-conditioned applicability.
- Selection uses a deterministic ε-lexicographic rule preserving cross-tier priority.
- No retraining of the predictor is required.
- Tested on Waymo Open Motion Dataset validation_interactive split (43,219 instances, K=6).
- Protocol B uses a 28-rule proxy catalog with oracle applicability.
- Safety+Legal violations reduced from 28.58% to 20.42%.
- Total violations reduced from 40.32% to 32.41%.
Entities
Institutions
- Waymo