PaperReading

nuPlan: A closed-loop ML-based planning benchmark for autonomous vehicles

January 2023

tl;dr: closed-loop benchmark for ML-based motion planners.

Overall impression

This paper is one pioneering paper/tech report on ML-based motion planner (neural planner). The report points out that in order to get a fair benchmark, closed loop simulation has to be used.

Background: AV perception has witnessed impressive progress. In contrast, existing solutions for AV planning are primarily based on carefully engineered expert systems. They require significant amount of work to adapt to new geographies and does not scale with training data. Having ML-based planning will pave the way to a full Software 2.0 stack.

Prediction focuses on the behavior of other agents, while planning relates to the ego vehicle behavior. Prediction is multimodal, and for each agent we predict the N most likely trajectories. Planning is unimodal.

Key ideas

Technical details

Notes