# Hidden Biases of End-to-End Driving Models

**Authors:** Bernhard Jaeger, Kashyap Chitta, Andreas Geiger  
**arXiv:** [2306.07957](https://arxiv.org/abs/2306.07957)  
**Date:** 2023-06-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.07957v1/

## Key points

- The paper explores the recent progress of end-to-end driving systems on CARLA
- Lateral recovery via a strong inductive bias towards target point following
- Longitudinal averaging of multimodal waypoint predictions for slowing down
- The authors investigate the drawbacks of these biases and identify principled alternatives
- They develop TF++, a simple end-to-end method that incorporates their insights and ranks first on the Longest6 and LAV benchmarks, gaining 14 driving score over the best prior work on Longest6
- This research highlights the importance of understanding biases in machine learning models and developing principled alternatives to improve performance
- The findings have significant implications for autonomous driving technology as they provide insights into how to design more effective end-to-end driving systems that can navigate complex environments with greater accuracy and safety.

## Summary

The paper "Hidden Biases of End-to-End Driving Models" by Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger explores the recent rapid progress of end-to-end driving systems on CARLA and identifies two biases that are critical for this progress. The authors note that these systems introduce changes to minor system components, making it unclear where the improvements come from. The first bias identified is lateral recovery via a strong inductive bias towards target point following, which is present in nearly all state-of-the-art methods. The second bias is longitudinal averaging of multimodal waypoint predictions for slowing down. The authors investigate the drawbacks of these biases and identify principled alternatives. They develop TF++, a simple end-to-end method that incorporates their insights and ranks first on the Longest6 and LAV benchmarks, gaining 14 driving score over the best prior work on Longest6. The findings of this paper highlight the importance of understanding biases in machine learning models and developing principled alternatives to improve performance. This research has significant implications for autonomous driving technology as it provides insights into how to design more effective end-to-end driving systems that can navigate complex environments with greater accuracy and safety.