[论文解读] Parting with Misconceptions about Learning-based Vehicle Motion Planning

Daniel Dauner, Marcel Hallgarten+2 more|arXiv (Cornell University)|Jun 13, 2023

Autonomous Vehicle Technology and Safety被引用 15

一句话总结

The paper shows open-loop ego-forecasting and closed-loop planning are misaligned, highlights the strong performance of a simple centerline-based baseline and rule-based IDM, and introduces PDM-Hybrid, a lightweight IDM-based planner with learned enhancements that achieves state-of-the-art results in nuPlan.

ABSTRACT

The release of nuPlan marks a new era in vehicle motion planning research, offering the first large-scale real-world dataset and evaluation schemes requiring both precise short-term planning and long-horizon ego-forecasting. Existing systems struggle to simultaneously meet both requirements. Indeed, we find that these tasks are fundamentally misaligned and should be addressed independently. We further assess the current state of closed-loop planning in the field, revealing the limitations of learning-based methods in complex real-world scenarios and the value of simple rule-based priors such as centerline selection through lane graph search algorithms. More surprisingly, for the open-loop sub-task, we observe that the best results are achieved when using only this centerline as scene context (i.e., ignoring all information regarding the map and other agents). Combining these insights, we propose an extremely simple and efficient planner which outperforms an extensive set of competitors, winning the nuPlan planning challenge 2023.

研究动机与目标

• Identify fundamental misalignments between open-loop ego-forecasting and closed-loop planning in nuPlan.
• Evaluate simple rule-based baselines versus learned planners on open- and closed-loop metrics.
• Propose a lightweight centerline-based baseline and a hybrid planner combining rule-based and learned components.
• Demonstrate that a centerline-only context can yield strong open-loop results and that rule-based methods can outperform learned models on closed-loop tasks.

提出的方法

• Analyze nuPlan dataset and benchmarking setup to compare open-loop and closed-loop evaluations.
• Develop PDM-Open: a centerline-conditioned MLP that predicts 8-second future waypoints using only centerline and ego history.
• Extend IDM with MPC-inspired elements to create PDM-Closed: multiple IDM-based trajectory proposals, simulation, and scoring to select the best trajectory.
• Create PDM-Hybrid: add learned offsets to long-horizon waypoints produced by PDM-Closed to improve long-horizon accuracy while preserving short-horizon reliability.
• Evaluate a variety of baselines (PlanCNN, Urban Driver, GC-PGP, IDM) and compare against nuPlan metrics (CLS-R, CLS-NR, OLS) and runtime.
• Provide an ablation study to understand the impact of correction horizon, proposals, and inputs on performance.

实验结果

研究问题

• RQ1 Do open-loop ego-forecasting metrics correlate with closed-loop driving performance in nuPlan?
• RQ2 Can a centerline-only representation suffice for strong open-loop performance, and how do rule-based methods fare in closed-loop tasks?
• RQ3 Does integrating a simple learned component with a rule-based planner (PDM-Hybrid) outperform purely learned or purely rule-based approaches on nuPlan?
• RQ4 What are the key factors (centerline extent, history, proposals, forecasting) that drive performance in a lightweight planning framework?

主要发现

• Open-loop and closed-loop evaluations are negatively correlated; strong ego-forecasting does not guarantee safe or effective closed-loop planning.
• A centerline-based rule-based IDM baseline can outperform state-of-the-art learned planners in closed-loop metrics on nuPlan.
• A centerline-only context yields competitive or superior open-loop performance, with PDM-Open achieving high OLS with minimal inputs.
• PDM-Closed improves closed-loop scores relative to IDM, and the PDM-Hybrid combines PDM-Closed with PDM-Open to yield strong performance across metrics.
• PDM-Hybrid achieves state-of-the-art results on nuPlan, and in the nuPlan Challenge 2023 outperformed many competitors in overall score when combining centerline planning with learned offsets.
• Ablation studies show the benefits of a 2-second correction horizon and the importance of environment forecasting and proposals for closed-loop performance.