# Parting with Misconceptions about Learning-based Vehicle Motion Planning

**Authors:** Daniel Dauner, Marcel Hallgarten, Andreas Geiger, Kashyap Chitta  
**arXiv:** [2306.07962](https://arxiv.org/abs/2306.07962)  
**Date:** 2023-06-13  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.07962v1/

## Key points

- Introduction of nuPlan as a groundbreaking advancement in vehicle motion planning research
- Emphasis on the need for precise short-term planning and long-horizon ego-forecasting in evaluation schemes
- Recommendation to address short-term planning and long-horizon ego-forecasting tasks independently
- Highlight of limitations of learning-based methods in complex real-world scenarios
- Value of simple rule-based priors like centerline selection through lane graph search algorithms
- Optimal results achieved by utilizing only the centerline as scene context for open-loop sub-task
- Proposal of an extremely simple and efficient planner that outperforms competitors
- Victory in the nuPlan planning challenge 2023 with the innovative approach

## Summary

The paper "Parting with Misconceptions about Learning-based Vehicle Motion Planning" by Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta introduces nuPlan as a groundbreaking advancement in vehicle motion planning research. This new era is marked by the release of a large-scale real-world dataset and evaluation schemes that demand precise short-term planning and long-horizon ego-forecasting. The authors highlight the challenge faced by existing systems in meeting both requirements simultaneously and conclude that these tasks should be addressed independently. The study delves into the current state of closed-loop planning within the field and sheds light on the limitations of learning-based methods when applied to complex real-world scenarios. It emphasizes the value of simple rule-based priors such as centerline selection through lane graph search algorithms. Surprisingly, the authors observe that for the open-loop sub-task, optimal results are achieved by utilizing only the centerline as scene context while disregarding information related to maps and other agents. Building on these insights, the authors propose an extremely simple and efficient planner that surpasses a wide range of competitors. Their innovative approach secures victory in the nuPlan planning challenge 2023. This comprehensive analysis challenges misconceptions surrounding learning-based vehicle motion planning and underscores the importance of leveraging rule-based priors for enhanced performance in real-world scenarios.