A behavior driven approach for sampling rare event situations for autonomous vehicles

Sarkar, Atrisha; Czamecki, Krzysztof · 2019 · Crossref

DOI: 10.1109/iros40897.2019.8967715

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 challenge of evaluating autonomous vehicle (AV) performance in rare event scenarios, such as crashes or near-misses, which are difficult to capture in naturalistic driving data due to their low probability. The authors argue that relying solely on field tests is prohibitively expensive, while standard simulation methods often lack realistic models of human driver behavior. To bridge this gap, the paper proposes a behavior-driven approach for rare event sampling based on the theory of bounded rationality. This method aims to estimate the probability of rare events and generate specific traffic situations that lead to them, offering greater interpretability than standard cross-entropy importance sampling techniques. The methodology models human driving behavior using a quantal response function extended for continuous actions and multi-objective utilities. The model incorporates three utility functions: safety (distance gap and time-to-collision) and progress (velocity). A rationality parameter vector controls the degree of sub-optimality for each utility, allowing the generation of diverse driving policies. The authors categorize these policies into eight distinct behavior types based on adherence to safety and progress constraints. To perform rare event sampling, they employ a Simulated Annealing optimization algorithm to identify the specific rationality parameters that maximize the probability of a rare event (defined as a vehicle cut-in scenario) within each behavior category. This optimized policy serves as the proposal distribution for importance sampling. The study evaluates the proposed approach using the University of Michigan SPMD naturalistic driving dataset, focusing on vehicle cut-in events. The authors compare their method against cross-entropy based importance sampling and crude Monte Carlo sampling. Results indicate that the bounded rationality approach provides a lower variance estimate of rare event probabilities compared to cross-entropy methods. Specifically, the proposed method achieves a 33% speed-up over cross-entropy sampling and a speed-up on the order of $10^4$ compared to crude Monte Carlo sampling. Furthermore, the model successfully fits the naturalistic dataset and generates new, realistic traffic situations that reflect the specific behavioral conditions leading to rare events. The significance of this work lies in its ability to provide both accurate probability estimates and interpretable behavioral insights for AV verification. By linking rare events to specific categories of human driving behavior (e.g., aggressive progress vs. safety adherence), the method aids in understanding the root causes of critical scenarios. This approach offers a more efficient and informative alternative to existing sampling techniques, facilitating better testing and validation of autonomous driving systems in complex, stochastic traffic environments.

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.

• situational awareness

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: dataset resource
• Theoretical Contribution: computational model