Abstract

An open-source modular benchmark for diffusion-based motion planners within production AD stacks. The monolithic ONNX graph (≈18,000 nodes) is decomposed, DPM-Solver++ is reimplemented in C++, and the components are integrated with ROS 2 / Autoware. Controlled component swaps reveal which design choices actually move closed-loop score and report a latency–CL-score Pareto frontier.

Accepted, ITSC 2026, Naples, Italy.