SCRP: Repurposing Trajectory Prediction for Safety-Critical Relevance Planning with Agent Selection

Trajectory planning for autonomous driving requires reasoning about interactions among multiple agents under real-time constraints. Existing approaches often rely on increased model complexity and iterative trajectory generation to achieve strong performance, but incur high computational cost and limit scalability. In contrast, one-shot planners generate trajectories in a single forward pass and improve efficiency, yet fail to match iterative methods in complex scenarios. We identify that this performance gap arises from the absence of decision-aware interaction reasoning rather than limitations in model capacity or inference style. This paper presents a safety-critical relevance planning (SCRP) framework that enables efficient and high-performance one-shot planning. SCRP builds upon a generic state encoder–interaction encoder–trajectory decoder backbone inherited from trajectory prediction models and enables decision-aware interaction reasoning through two components: perceived safety modeling and risk-conditioned adaptive agent selection. A perceived safety modeling mechanism learns temporally evolving interaction risk and distills future-supervised safety cues into the latent interaction representation. In addition, a risk-conditioned adaptive selection mechanism allocates a scene-dependent interaction cardinality by retaining the most relevant agents under a risk-aware residual-mass criterion. The resulting interaction set adapts its capacity to scene complexity while suppressing weakly relevant context and preserving computational efficiency. Extensive experiments on large-scale autonomous driving benchmarks, including Waymo and nuPlan, demonstrate that SCRP consistently outperforms state-of-the-art (SOTA) planning models, particularly in challenging interactive and safety-critical scenarios. Adding SCRP reduces inference latency by 4.5%–9.1% relative to the corresponding prediction backbones, improving planning quality while retaining an efficient one-shot inference pipeline.