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

Rohan Chandra, Mingyu Wang, Mac Schwager, Dinesh Manocha

TL;DR

Autonomous vehicles (AVs) often struggle to navigate human traffic because they can’t "read the room." This paper introduces a framework that uses graph theory to quantify human driver aggressiveness (the CMetric) and maps it to a mathematical risk-sensitivity parameter. By plugging these parameters into a game-theoretic planner, AVs can anticipate and safely interact with both "road ragers" and cautious drivers, reducing navigation error by 10% compared to standard methods.

What Problem Does This Paper Address?

Driving is a social game. Human drivers exhibit a spectrum of behaviors, from aggressive (speeding, frequent lane changes, tight cut-ins) to conservative (driving slowly, staying in the right lane). Most current AV planners either assume a "one-size-fits-all" neutral behavior for humans or require massive amounts of training data to classify intent.

When an AV assumes everyone is "neutral," it fails in two ways:

  1. Safety: It might get too close to an aggressive driver who is likely to make sudden moves.
  2. Efficiency: It might get stuck in a "deadlock" with a conservative driver, where both vehicles wait for the other to move first.

The challenge is to create a computationally tractable way to estimate a human's internal risk tolerance in real-time and use that information to plan a trajectory that is both safe and socially aware.

Methods

The authors propose a four-step pipeline:

  1. CMetric Calculation: Instead of using deep learning, they use a graph-theoretic model called CMetric. They represent the traffic scene as a weighted undirected graph where vehicles are nodes and the distances between them are edges. By calculating "closeness" and "degree centrality," they can objectively score a driver's aggressiveness based on their trajectory over a 5-second window.
  2. Mapping to Risk Sensitivity: They bridge the gap between behavior and math by using linear regression to map the CMetric score to an entropic risk parameter ($\theta$). This parameter defines whether an agent is risk-seeking ($\theta < 0$), risk-neutral ($\theta = 0$), or risk-averse ($\theta > 0$).
  3. Game-Theoretic Planning: The AV treats the interaction as a dynamic game. It seeks a Nash Equilibrium—a state where no player can improve their outcome by changing their strategy alone. They use a modified version of a risk-sensitive planner that solves a set of Linear-Exponential-Quadratic (LEQ) equations iteratively.
  4. Receding Horizon Control: The AV predicts the human's future motion based on their identified risk profile and executes its own path in a continuous loop.

Datasets

The study primarily utilizes the OpenAI Gym highway-env simulator, a common benchmark for autonomous driving and tactical decision-making.

  • Training Data: Generated by simulating agents with a fixed array of risk sensitivity values (from -5.0 to +5.0) to build the linear mapping between CMetric and $\theta$.
  • Test Data: Generated by human participants driving in the simulator using a keyboard to capture "natural" aggressive and conservative trajectories.

Key Insights and Ideas

  • Aggressiveness is Risk-Seeking: The core "aha" moment is the formal link between social behavior (aggressiveness) and mathematical risk (entropic risk). An aggressive driver is essentially a "risk-seeking" agent in the LEQ framework.
  • Emergent Social Behavior: The planner doesn't need explicit rules like "yield to the fast car." Instead, the behavior emerges from the math. When the AV identifies a risk-seeking human, the optimal mathematical solution for the AV (if it is risk-averse) is to increase its distance and yield, mimicking how a defensive human driver would act.
  • Lane Changing as a Proxy: The authors found a strong correlation between risk-seeking behavior ($\theta < 0$) and the frequency of lane changes, providing a secondary way to validate their risk models.

Key Results

  • Safety Improvement: Modeling human risk profiles reduced the error (discrepancy between expected and actual safe distance) by 10% compared to baselines that assume neutral human behavior.
  • Human-Like Trajectories: In a user study with 27 participants, 96% (26 out of 27) were able to correctly distinguish between aggressive and conservative drivers based solely on the trajectories generated by the planner.
  • Deadlock Resolution: The model successfully demonstrated the ability to avoid "brief deadlocks" with conservative drivers by recognizing their risk-aversion and taking the initiative to merge first.

Limitations and Future Work

  • Open-Loop Constraints: The current evaluation is largely "open-loop," meaning the human behavior model and the trajectory planner were tested in somewhat separate stages. The authors aim to move toward a closed-loop simulator where the human and AV react to each other's changes in real-time.
  • Simulator Complexity: The authors noted that keeping the environment configurations identical across different simulators (one for behavior modeling, one for planning) was cumbersome.
  • Real-World Noise: While CMetric is data-driven, it hasn't been tested against the high-sensor-noise environments of real-world urban driving. Future work involves testing the robustness of the graph-theoretic measures against occlusions and sensor errors.

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},
}

Loading related content...

Figure