Game-Theoretic Planning for Autonomous Driving among Risk-Aware Human Drivers

Rohan Chandra, Mingyu Wang, Mac Schwager, Dinesh Manocha

Abstract

We present a novel approach for risk-aware planning with human agents in multi-agent traffic scenarios. Our approach takes into account the wide range of human driver behaviors on the road, from aggressive maneuvers like speeding and overtaking, to conservative traits like driving slowly and conforming to the right-most lane. In our approach, we learn a mapping from a data-driven human driver behavior model called the CMetric to a driver's entropic risk preference. We then use the derived risk preference within a game-theoretic risk-sensitive planner to model risk-aware interactions among human drivers and an autonomous vehicle in various traffic scenarios. We demonstrate our method in a merging scenario, where our results show that the final trajectories obtained from the risk-aware planner generate desirable emergent behaviors. Particularly, our planner recognizes aggressive human drivers and yields to them while maintaining a greater distance from them. In a user study, participants were able to distinguish between aggressive and conservative simulated drivers based on trajectories generated from our risk-sensitive planner. We also observe that aggressive human driving results in more frequent lane-changing in the planner. Finally, we compare the performance of our modified risk-aware planner with existing methods and show that modeling human driver behavior leads to safer navigation.

Introduction

Autonomous vehicles (AVs) often struggle to navigate human traffic because they lack the ability to "read the room." Human drivers exhibit a broad spectrum of behaviors, ranging from aggressive (speeding, frequent lane changes) to conservative (driving slowly, staying in the right lane). Most current AV planners either assume a uniform "neutral" behavior for all humans or require massive datasets to classify intent. When an AV assumes everyone is neutral, it risks safety by getting too close to aggressive drivers or efficiency by entering deadlocks with conservative ones. This paper introduces a framework that quantifies human driver aggressiveness using graph theory and maps it to a mathematical risk-sensitivity parameter, enabling safer and more socially aware navigation.

Methodology

The proposed framework utilizes a four-step pipeline to integrate human behavior into trajectory planning:

  1. CMetric Calculation: Traffic scenes are represented as weighted undirected graphs where vehicles are nodes and distances are edges. The CMetric uses "closeness" and "degree centrality" to objectively score aggressiveness based on a 5-second trajectory window.
  2. Mapping to Risk Sensitivity: Linear regression maps the CMetric score to an entropic risk parameter ($\theta$). This parameter categorizes agents as risk-seeking ($\theta < 0$), risk-neutral ($\theta = 0$), or risk-averse ($\theta > 0$).
  3. Game-Theoretic Planning: The interaction is treated as a dynamic game seeking a Nash Equilibrium. A modified risk-sensitive planner solves Linear-Exponential-Quadratic (LEQ) equations iteratively.
  4. Receding Horizon Control: The AV predicts future motion based on the identified risk profile and executes its path in a continuous loop.
The four-step pipeline for risk-aware planning, from trajectory input to risk-aware execution.

Experimental Setup

The study utilized the OpenAI Gym highway-env simulator. Training data was generated by simulating agents with fixed risk sensitivity values ($-5.0$ to $+5.0$) to establish the linear mapping between CMetric and $\theta$. Test data involved human participants driving in the simulator to capture natural aggressive and conservative trajectories.

Results and Discussion

The core finding of this research is the formal link between social behavior and mathematical risk. An aggressive driver is effectively a "risk-seeking" agent in the LEQ framework. By identifying these profiles, desirable social behaviors emerge mathematically without explicit rules. For instance, when the AV identifies a risk-seeking human, the optimal solution is to increase distance and yield.

A merging scenario where the planner identifies the red car as conservative and the blue car as aggressive, resulting in the conservative agent yielding.

Key quantitative and qualitative results include:

  • Safety Improvement: Modeling human risk profiles reduced navigation error by 10% compared to neutral-assumption baselines.
  • Human-Like Trajectories: In a study with 27 participants, 96% correctly distinguished between aggressive and conservative drivers based on the planner's generated trajectories.
  • Deadlock Resolution: The model successfully avoided deadlocks with conservative drivers by recognizing their risk-aversion and taking the initiative to merge first.
  • Behavioral Proxies: A strong correlation was observed between risk-seeking behavior ($\theta < 0$) and the frequency of lane changes.

Conclusion and Future Work

While the framework demonstrates significant improvements in safety and social awareness, current evaluations are largely open-loop. Future work will focus on closed-loop simulations where humans and AVs react to each other in real-time. Additionally, the researchers aim to test the robustness of graph-theoretic measures like CMetric against real-world sensor noise and occlusions in urban environments.

Summary updated

Abstract

We present a novel approach for risk-aware planning with human agents in multi-agent traffic scenarios. Our approach takes into account the wide range of human driver behaviors on the road, from aggressive maneuvers like speeding and overtaking, to conservative traits like driving slowly and conforming to the right-most lane. In our approach, we learn a mapping from a data-driven human driver behavior model called the CMetric to a driver's entropic risk preference. We then use the derived risk preference within a game-theoretic risk-sensitive planner to model risk-aware interactions among human drivers and an autonomous vehicle in various traffic scenarios. We demonstrate our method in a merging scenario, where our results show that the final trajectories obtained from the risk-aware planner generate desirable emergent behaviors. Particularly, our planner recognizes aggressive human drivers and yields to them while maintaining a greater distance from them. In a user study, participants were able to distinguish between aggressive and conservative simulated drivers based on trajectories generated from our risk-sensitive planner. We also observe that aggressive human driving results in more frequent lane-changing in the planner. Finally, we compare the performance of our modified risk-aware planner with existing methods and show that modeling human driver behavior leads to safer navigation.
BibTeX
@inproceedings{chandra2022game,
  author = {Rohan Chandra and Mingyu Wang and Mac Schwager and Dinesh Manocha},
  title = {{Game-Theoretic Planning for Autonomous Driving among Risk-Aware Human Drivers}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2022},
  doi = {10.1109/icra46639.2022.9811865},
}
Figure