Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Integrating End-to-End and Modular Driving Approaches for Online Corner Case Detection in Autonomous Driving

Gemb Kaljavesi, Xiyan Su, Frank Diermeyer

TL;DR

The paper addresses online corner-case detection for autonomous driving by integrating modular and end-to-end approaches. It proposes a disagreement-based detector where a modular primary system operates in parallel with an end-to-end secondary network, leveraging divergences in lateral trajectories and hazard-oriented longitudinal signals, with Lat(t) and Long(t) defined to quantify discrepancy. The method is implemented on a real vehicle (Autoware Universe and tf++ on the EDGAR platform) and trained through sim-to-real transfer in CARLA followed by real-data fine-tuning. Qualitative results indicate potential, particularly in longitudinal hazard detection, while highlighting challenges from sim-to-real gaps and limited scenario diversity, thereby motivating quantitative evaluation and expanded datasets in future work.

Abstract

Online corner case detection is crucial for ensuring safety in autonomous driving vehicles. Current autonomous driving approaches can be categorized into modular approaches and end-to-end approaches. To leverage the advantages of both, we propose a method for online corner case detection that integrates an end-to-end approach into a modular system. The modular system takes over the primary driving task and the end-to-end network runs in parallel as a secondary one, the disagreement between the systems is then used for corner case detection. We implement this method on a real vehicle and evaluate it qualitatively. Our results demonstrate that end-to-end networks, known for their superior situational awareness, as secondary driving systems, can effectively contribute to corner case detection. These findings suggest that such an approach holds potential for enhancing the safety of autonomous vehicles.

Integrating End-to-End and Modular Driving Approaches for Online Corner Case Detection in Autonomous Driving

TL;DR

The paper addresses online corner-case detection for autonomous driving by integrating modular and end-to-end approaches. It proposes a disagreement-based detector where a modular primary system operates in parallel with an end-to-end secondary network, leveraging divergences in lateral trajectories and hazard-oriented longitudinal signals, with Lat(t) and Long(t) defined to quantify discrepancy. The method is implemented on a real vehicle (Autoware Universe and tf++ on the EDGAR platform) and trained through sim-to-real transfer in CARLA followed by real-data fine-tuning. Qualitative results indicate potential, particularly in longitudinal hazard detection, while highlighting challenges from sim-to-real gaps and limited scenario diversity, thereby motivating quantitative evaluation and expanded datasets in future work.

Abstract

Online corner case detection is crucial for ensuring safety in autonomous driving vehicles. Current autonomous driving approaches can be categorized into modular approaches and end-to-end approaches. To leverage the advantages of both, we propose a method for online corner case detection that integrates an end-to-end approach into a modular system. The modular system takes over the primary driving task and the end-to-end network runs in parallel as a secondary one, the disagreement between the systems is then used for corner case detection. We implement this method on a real vehicle and evaluate it qualitatively. Our results demonstrate that end-to-end networks, known for their superior situational awareness, as secondary driving systems, can effectively contribute to corner case detection. These findings suggest that such an approach holds potential for enhancing the safety of autonomous vehicles.
Paper Structure (14 sections, 4 equations, 4 figures, 1 table)

This paper contains 14 sections, 4 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Contribution
  4. Proposed Online Corner Case Detection
  5. Experiment Design
  6. Training
  7. Deployment
  8. Corner Case Detection
  9. Evaluation
  10. Results
  11. Lateral
  12. Longitudinal
  13. Discussion
  14. Conclusion

Figures (4)

  • Figure 1: Overview of the proposed method for online corner case detection through the integration of end-to-end and modular software.
  • Figure 2: Overview of the experiment design. The upper box illustrates the training process, including the inputs and outputs in the simulation and with real data. The box in the bottom left depicts the implementation of our method, showing only the modules relevant to the method. The test track is shown in the bottom right.
  • Figure 3: Visualization of the lateral part of the trajectory of both the modular system and the end-to-end system over time, during an overtaking maneuver, and the distance between the two. The corresponding scenario is additionally illustrated below.
  • Figure 4: Visualization of the predicted speed class in orange and the target speed specified by the modular system. The corner case in which the safety driver has braked is shown with the red dashed lines. The corresponding scenario is additionally illustrated below.