I Know You Can't See Me: Dynamic Occlusion-Aware Safety Validation of Strategic Planners for Autonomous Vehicles Using Hypergames

Kahn, Maximilian; Sarkar, Atrisha; Czarnecki, Krzysztof · 2022 · Crossref

DOI: 10.1109/icra46639.2022.9812041

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 safety validation for autonomous vehicle (AV) strategic planners in scenarios involving dynamic occlusion, where vehicles obstruct each other’s view. While static occlusions (e.g., buildings) are well-studied, dynamic occlusions are transient and unpredictable, creating significant risks because standard game-theoretic planners assume all agents share a common view of traffic. The authors identify a gap in existing safety validation frameworks, which often fail to account for the multi-agent nature of these risks or the breakdown of common-knowledge assumptions when occlusions occur. To address this, the authors propose a white-box, scenario-based safety validation framework grounded in hypergame theory. They introduce a novel Dynamic Occlusion Risk (DOR) measure that quantifies risk by comparing the outcomes of an "occlusion-resolved" game (where all vehicles see each other) against "occlusion-naive" games (where vehicles ignore occluded spaces). The method assumes a setting where the AV is occlusion-aware while human drivers are occlusion-naive. The framework utilizes a search-based approach to augment naturalistic driving data with synthetic occluding vehicles. It employs voxel-based raycasting to detect occlusions and injects vehicles into realistic configurations to generate critical test scenarios. The system then constructs level-1 hypergames to simulate how occlusion-naive drivers would plan, identifying scenarios where the divergence between the AV’s strategy and human drivers’ strategies leads to collisions. The approach was evaluated using the Waterloo Multi-Agent (WMA) database, which contains over 3,500 vehicles recorded at a busy intersection. The authors compared their accelerated validation method against direct validation on naturalistic data. The results demonstrated a 4,000% speedup in generating occlusion-causing crashes compared to relying solely on naturalistic data. The method achieved more diverse coverage of scenarios and successfully generated commonly observed dynamic occlusion crashes that were not present in the input dataset. This indicates the framework’s ability to generalize beyond observed data and automatically identify high-risk situations. The significance of this work lies in providing a scalable, automated method for validating strategic AV planners under realistic, high-risk conditions. By leveraging hypergames, the framework explicitly models the mismatch in situational awareness caused by dynamic occlusions, offering a more rigorous safety assessment than traditional methods. This contributes to the development of safer AVs by ensuring planners are robust against the transient and complex risks inherent in shared 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.

• visual occlusion