Abstract
As autonomous vehicles (AVs) and human-driven vehicles begin to coexist, integrating AVs into the existing traffic ecology without causing friction is a critical challenge. This paper presents a human-like decision-making framework for AVs that incorporates personalized driving styles—aggressive, normal, and conservative—by balancing safety, comfort, and efficiency. Using noncooperative game theory, specifically Nash equilibrium and Stackelberg game models, the framework enables AVs to interact socially with human drivers. Combined with potential field methods and model predictive control (MPC) for motion planning, the system provides a robust solution for complex maneuvers like merging and overtaking. Results demonstrate that the Stackelberg approach reduces decision-making costs by over 20% compared to Nash equilibrium under normal driving conditions.
Introduction: The "Misfit" Problem
The transition to autonomous driving faces a significant hurdle: the "misfit" between robotic, overly-cautious AVs and unpredictable human drivers. This discrepancy can lead to traffic inefficiency or even accidents. Furthermore, passengers have diverse needs; a commuter might prioritize speed, while others prioritize comfort. To address this, AVs must not only drive safely but also "negotiate" road space in a way that is recognizable and acceptable to humans.
Methodology
The researchers developed a three-layered architecture to bridge the gap between high-level intent and low-level control.
Human-Like Modeling
Using the NGSIM dataset (I-80 and US-101 freeway data), the team extracted characteristic parameters to define three driving styles: Aggressive, Normal, and Conservative. These styles are quantified by three key metrics: velocity (efficiency), acceleration (comfort), and time headway (safety).
Game Theoretic Decision-Making
The core of the framework treats lane-changing as a social negotiation between the Ego Car (EC) and Adjacent Cars (AC). Two models were evaluated:
- Nash Equilibrium: A simultaneous game where both vehicles act as equal players minimizing their own costs.
- Stackelberg Game: A leader-follower model where the EC (leader) anticipates the AC's (follower) response. This hierarchical approach better reflects the proactive nature of human lane-changing.
Motion Planning
High-level decisions are translated into trajectories using a combination of Artificial Potential Fields (APF) and Model Predictive Control (MPC). APF handles obstacle avoidance and goal attraction, while MPC ensures the resulting path is physically feasible for the vehicle.
Key Results and Insights
The study highlights that lane changing is fundamentally a social problem rather than just a geometric one.
- Superiority of Stackelberg: The Stackelberg approach proved more efficient, reducing decision-making costs by over 20% compared to the Nash approach in normal driving scenarios. By accounting for the other driver's likely reaction, the AV makes smoother, more decisive moves.
- Personalization via Cost Functions: Human-like behavior is achieved by adjusting weights in a cost function. For instance, the aggressive model weights travel efficiency at 80%, while the conservative model prioritizes safety at 70%.
- Validation: The framework was successfully tested in simulation for merging and overtaking. Aggressive AVs proactively created gaps, while conservative ones waited for larger safety buffers.
Conclusion and Future Directions
While the framework successfully demonstrates human-like interaction, challenges remain regarding computational complexity for real-time deployment. Future work will focus on optimizing the game-theoretic algorithms and expanding the system to handle complex urban environments involving intersections and pedestrians. The researchers plan to move from MATLAB simulations to Hardware-in-the-Loop (HiL) testing to further validate the system's real-world applicability.
Summary updated