Learning from Demonstrations of Critical Driving Behaviours Using Driver's Risk Field

Du, Yurui; Acerbo, Flavia Sofia; Kober, Jens; Son, Tong Duy · 2023 · Crossref

DOI: 10.1016/j.ifacol.2023.10.1376

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the limitations of imitation learning (IL) in autonomous vehicle (AV) planning, specifically regarding sample inefficiency and poor generalization in safety-critical scenarios. While IL is widely used in industry, previous approaches often reach a performance plateau where adding more training data fails to improve the policy, largely because such methods are rarely tested on critical edge cases. The authors propose a method to enhance both safety and training efficiency by leveraging a Driver’s Risk Field (DRF), a parametric model of human driving behavior implemented in a multi-agent traffic simulator based on the Lyft Prediction Dataset. The methodology involves two main components. First, the authors present an IL model that utilizes spline coefficient parameterization and offline expert queries to improve training efficiency. Second, to address the weakness of the learned policy in critical situations, they synthetically generate critical driving scenarios. This is achieved through the optimization of parameters within the DRF, which allows for the creation of specific, high-risk situations that expose policy vulnerabilities. The IL model is then retrained on this augmented data. The approach relies on the expressivity and interpretability of the DRF to continuously improve the learned policy by targeting areas where the model previously failed or underperformed. The findings demonstrate that this approach effectively overcomes the performance plateau typical of standard IL methods. By exposing the weaknesses of the learned policy through synthetically generated critical scenarios, the method allows for targeted improvement. The use of the DRF enables the generation of relevant, safety-critical data that might be scarce in real-world datasets. Consequently, the retrained IL model shows enhanced performance in these critical scenarios, validating the effectiveness of using parametric risk models to guide data augmentation and policy refinement. The significance of this work lies in its contribution to safer and more efficient autonomous driving systems. It provides a framework for moving beyond passive data collection by actively generating challenging scenarios based on a interpretable model of human risk perception. This addresses a critical gap in AV development, where safety-critical generalization is paramount. The integration of the Driver’s Risk Field with imitation learning offers a pathway to robust policies that can handle rare but dangerous situations, thereby improving the overall reliability of autonomous planning modules.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction
• situational awareness
• in vehicle coaching

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software
• Theoretical Contribution: computational model, theory or model