# Rethinking Imitation-based Planner for Autonomous Driving

**Authors:** Jie Cheng, Yingbing Chen, Xiaodong Mei, Bowen Yang, Bo Li, Ming Liu  
**arXiv:** [2309.10443](https://arxiv.org/abs/2309.10443)  
**Date:** 2023-09-19  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2309.10443v1/

## Key points

- The study focuses on the effectiveness of imitation-based driving planners in recent years
- Introduces nuPlan to provide a standardized benchmark for fair comparisons
- Explores essential features for ego planning and data augmentation techniques
- Identifies an imitation gap overlooked by current learning systems
- Proposes the PlanTF model as a strong baseline model with competitive performance
- Demonstrates that purely imitation-based planners can outperform methods involving hand-crafted rules
- Highlights superior generalization capabilities of PlanTF in long-tail cases compared to rule-based methods
- Reveals shortcut learning issues related to historical motions and single-frame kinematic states
- Shows that planning based solely on the autonomous vehicle's current position leads to superior closed-loop performance
- Introduces an attention-based state dropout encoder (SDE) as a solution to mitigate shortcut learning problems within imitation-based planners

## Summary

The study "Rethinking Imitation-based Planner for Autonomous Driving" by Jie Cheng, Yingbing Chen, Xiaodong Mei, Bowen Yang, Bo Li, and Ming Liu delves into the effectiveness of imitation-based driving planners in recent years. The absence of a standardized benchmark has made it challenging to compare various designs accurately. To address this issue, the researchers introduce nuPlan, which provides a large-scale real-world dataset and a standardized closed-loop benchmark for fair comparisons. Through their comprehensive study utilizing nuPlan, the researchers explore two critical yet underexplored aspects of imitation-based planners: identifying essential features for ego planning and effective data augmentation techniques to reduce compounding errors. They also shed light on an imitation gap overlooked by current learning systems. By integrating their findings, they propose a strong baseline model called PlanTF. The results of their study demonstrate that a well-designed purely imitation-based planner can achieve highly competitive performance compared to state-of-the-art methods involving hand-crafted rules. Additionally, the PlanTF model exhibits superior generalization capabilities in long-tail cases. The researchers highlight that while rule-based methods perform well in ordinary scenarios, they struggle to generalize in long-tail situations. In contrast, PlanTF shows stronger generalization capabilities. In conclusion,<kg> this study systematically examines crucial design aspects of imitation-based planners using the nuPlan benchmark.</kg> The findings reveal that catastrophic shortcut learning often occurs for input features like historical motions and single-frame kinematic states.<kg> Surprisingly,</kg> planning solely based on the autonomous vehicle's current position leads to superior closed-loop performance.<kg> To address shortcut learning issues effectively,</kg> the researchers introduce an attention-based state dropout encoder (SDE). This SDE proves to be an effective solution to mitigate shortcut learning problems within imitation-based planners for autonomous driving applications. Overall, this research contributes valuable insights into improving imitation-based planners for autonomous driving systems and offers a publicly available project website for further exploration and implementation of their models and benchmarks.