PDMScorer implements the core closed-loop scoring logic for trajectory proposals, computing sub-scores for collision, drivable area, traffic light compliance, progress, time-to-collision, lane keeping, and comfort. It is configurable via weights and thresholds, and can be reused independently of the PDM planner for evaluating any trajectory set against a scenario.

ConditionalUnet1D is a 1D conditional U-Net architecture for diffusion-based trajectory generation. It supports variable-length conditioning and denoising steps, and is designed to be pluggable into any diffusion-based planning agent. The implementation is self-contained and demonstrates non-obvious skill in adapting diffusion models to trajectory domains.

BatchIDMPolicy implements an Intelligent Driver Model (IDM) policy that generates multiple trajectory proposals in batch for a given ego state and surrounding agents. It is optimized for vectorized operations and can be extracted as a standalone module for any planning system requiring physics-based trajectory candidates.

TransfuserBackbone is a multi-scale fusion transformer that processes camera images and LiDAR point clouds to produce a shared feature representation for planning. It implements cross-attention between modalities and can be reused as a sensor fusion backbone in other autonomous driving models, independent of the Transfuser agent.