# Imitation Is Not Enough: Robustifying Imitation with Reinforcement Learning for Challenging Driving Scenarios

**Authors:** Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Becca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, Dragomir Anguelov, Sergey Levine  
**arXiv:** [2212.11419](https://arxiv.org/abs/2212.11419)  
**Date:** 2022-12-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.11419v1/

## Key points

- Authors explore combining imitation learning (IL) and reinforcement learning for autonomous driving
- First application of combined IL and RL techniques in autonomous driving using real-world human driving data
- Trained policy using over 100k miles of urban driving data
- Evaluated performance in test scenarios with different collision risk levels
- Integration of reinforcement learning with simple rewards led to significant improvements in policy safety and reliability compared to IL alone
- Potential of integrated methodology to enhance autonomous vehicle systems' ability to navigate challenging scenarios effectively, prioritizing safety and reliability

## Summary

In the paper titled "Imitation Is Not Enough: Robustifying Imitation with Reinforcement Learning for Challenging Driving Scenarios," authors Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Becca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, Dragomir Anguelov, and Sergey Levine explore the combination of imitation learning (IL) and reinforcement learning to enhance driving policies in autonomous vehicles. This novel approach marks the first application of combined imitation and reinforcement learning techniques in autonomous driving that leverages substantial amounts of real-world human driving data. The study involves training a policy using over 100k miles of urban driving data and evaluating its performance in test scenarios categorized by different levels of collision risk. By integrating reinforcement learning with simple rewards into the training process, the authors demonstrate significant improvements in policy safety and reliability compared to those solely based on imitation learning. The findings highlight the potential of this integrated methodology to enhance autonomous vehicle systems' ability to navigate challenging driving scenarios effectively while prioritizing safety and reliability.