Abstract
In this article, we propose an optimization-based integrated behavior planning and motion control scheme, which is an interpretable and adaptable urban autonomous driving solution that complies with complex traffic rules while ensuring driving safety. Inherently, to ensure compliance with traffic rules, an innovative design of potential functions (PFs) is presented to characterize various traffic rules related to traffic lights, traversable and non-traversable traffic line markings, etc. These PFs are further incorporated as part of the model predictive control (MPC) formulation. In this sense, high-level behavior planning is attained implicitly along with motion control as an integrated architecture, facilitating flexible maneuvers with safety guarantees. As demonstrated from a series of simulations with challenging scenarios in CARLA, it is noteworthy that the proposed framework admits real-time performance and high generalizability.
Introduction
Autonomous driving systems traditionally rely on a "stacked" architecture where perception, behavior planning, motion planning, and control are treated as separate, sequential modules. While modular, this approach often introduces a trade-off between computational efficiency and interpretability. Heuristic-based planners can be overly rigid in dynamic urban environments, while learning-based methods like Reinforcement Learning often lack the transparency required for safety-critical systems.
This research addresses these challenges by merging high-level behavior planning and low-level motion control into a single mathematical optimization problem. By treating traffic rules as continuous "potential fields," the system enables autonomous vehicles (AVs) to make complex behavioral decisions—such as when to overtake—implicitly through a unified Model Predictive Control (MPC) framework.
Methodology
Integrated Architecture
The core of the proposed system is an Integrated Behavior Planning and Motion Control Framework based on Nonlinear Model Predictive Control (NMPC). This architecture combines vehicle goals (waypoint tracking, energy efficiency, passenger comfort) with safety and regulatory constraints (collision avoidance, traffic rules) into a single cost function.
Innovative Potential Functions (PFs)
To represent the "discrete" nature of traffic rules in a continuous optimization space, the researchers developed specialized Artificial Potential Fields (APFs). These functions act as virtual "forces" that guide the vehicle:
- Non-traversable lines (solid): Modeled with high repulsive forces to prevent crossing.
- Traversable lines (dashed): Modeled as a "valley" shape. This keeps the vehicle centered under normal conditions but allows it to "climb" over the barrier to overtake if the objective function determines that maintaining speed is more valuable than staying in the lane.
- Traffic Lights: A novel PF restricts both longitudinal and lateral movement during red lights, ensuring smooth deceleration behind the stop line.
Optimization and Real-Time Performance
The framework utilizes a high-fidelity bicycle model to ensure planned trajectories are physically achievable. The resulting nonlinear programming (NLP) problem is solved using the IPOPT solver within the CasADi framework. By avoiding Mixed-Integer Programming, which is computationally expensive for discrete decisions, the system achieves real-time control loops.
Experimental Evaluation
The framework was rigorously tested in the CARLA simulator across three challenging urban scenarios: navigating complex roundabouts, multi-lane adaptive cruise control (ACC) with overtaking, and intersection driving with dynamic traffic lights.
Key performance metrics include:
- Success Rates: The system achieved a 93.3% success rate in roundabouts and 80-84% in other urban scenarios.
- Computational Efficiency: Average computation times ranged from 6.8ms to 11ms, well within the 50ms requirement for 20Hz real-time control.
- Precision: Position tracking error remained low, typically between 0.42m and 0.49m, even in dense traffic.
Discussion and Future Work
A significant insight from this work is the emergence of implicit behavior. Overtaking maneuvers are not explicitly programmed; they occur naturally when the mathematical "cost" of slowing down behind a lead vehicle exceeds the cost of crossing a traversable lane marking. This provides a level of flexibility and smoothness often missing in heuristic systems.
However, limitations remain. The current system assumes surrounding vehicles will maintain their current paths (naive prediction), which can lead to issues with aggressive agents. Future research will focus on integrating sophisticated trajectory prediction and developing "guiding PFs" for unmarked zones, such as large open intersections.
Summary updated