MUSE: Target-Guided Multimodal Scoring Ensemble for Safe Autonomous Planning

Abstract

Imitation learning planners can behave unsafely under distribution shift, since long-tailed driving data contain few safety-critical events. A common remedy is to generate multimodal candidate trajectories and select them with a hybrid evaluator, but existing methods often suffer from mode collapse and brittle hard constraints. We propose MUSE: Target-Guided Multimodal Scoring Ensemble for Safe Autonomous Planning. MUSE anchors multimodal generation on a compact, spatially diverse set of navigational target candidates selected by a learned target scorer with Route-Constrained Distance-Based Top-K selection, improving behavioral coverage and mitigating mode collapse; candidate trajectories are then generated conditioned on these targets. Candidates are ranked by a hybrid evaluator that fuses a learned long-horizon value estimate with short-horizon rule-based safety constraints and a soft risk metric for safety assurance. Extensive closed-loop experiments on the nuPlan benchmark show substantial gains in safety and overall driving performance: on Test14, MUSE achieves a collision score of 99.62 and NR-CLS of 93.17; on Test14-hard, it achieves a collision score of 98.51 and NR-CLS of 81.88.