# Safe Occlusion-aware Autonomous Driving via Game-Theoretic Active Perception

**Authors:** Zixu Zhang, Jaime F. Fisac  
**arXiv:** [2105.08169](https://arxiv.org/abs/2105.08169)  
**Date:** 2021-05-17  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2105.08169v1/

## Key points

- Novel approach for safe trajectory planning in autonomous driving systems
- Addresses challenge of reasoning about potential hazards beyond field-of-view due to occlusions
- Proposes hybrid zero-sum dynamic game framework between evader (autonomous vehicle) and initially hidden pursuer (traffic participant)
- Analysis of game yields optimal strategies for both players and set of initial conditions for collision avoidance
- Trajectory planning framework provides worst-case safety guarantees while minimizing conservativeness by actively avoiding other road users as soon as they are detected
- Framework is agnostic to driving environment and suitable for various motion planners
- Algorithm tested on challenging urban and highway driving scenarios using CARLA simulator
- Experimental results show successful navigation and collision avoidance in different interaction scenarios
- Introduces new analytical approach for safe trajectory planning under occlusions in autonomous driving systems
- Optimal strategies obtained along with worst-case safety guarantees

## Summary

This paper presents a novel approach for safe trajectory planning in autonomous driving systems, specifically addressing the challenge of reasoning about potential hazards beyond the field-of-view due to occlusions. The authors propose a hybrid zero-sum dynamic game framework between the autonomous vehicle (evader) and an initially hidden traffic participant (pursuer). The pursuer's state is initially unknown to the evader but may later be discovered by the vehicle's sensors. The analysis of this game yields optimal strategies for both players as well as a set of initial conditions from which the autonomous vehicle is guaranteed to avoid collisions. This theoretical result is then leveraged to develop a trajectory planning framework that provides worst-case safety guarantees while minimizing conservativeness by actively avoiding other road users as soon as they are detected. The proposed framework is agnostic to the driving environment and suitable for various motion planners. To demonstrate its effectiveness, the algorithm is tested on challenging urban and highway driving scenarios using the CARLA simulator. Experimental results show successful navigation and collision avoidance in different interaction scenarios, such as approaching intersections with pursuers outside or inside danger zones, overtaking trucks while avoiding oncoming traffic, and more. Overall, this paper introduces a new analytical approach for safe trajectory planning under occlusions in autonomous driving systems. By formulating it as a dynamic game between the evader and pursuer, optimal strategies are obtained along with worst-case safety guarantees. The experimental results highlight the effectiveness of this approach in various challenging driving scenarios.