Abstract Navigating complex urban environments requires autonomous vehicles (AVs) to balance high-level strategic reasoning with precise, safe motion planning. This paper introduces FFStreams++, a hybrid framework that bridges this gap. By integrating a deep-learning-based trajectory predictor (QCNet) with a symbolic PDDL-based heuristic planner, FFStreams++ enables AVs to perform sophisticated maneuvers like unprotected left turns and high-speed overtaking. Experimental results on the CommonRoad benchmark demonstrate that FFStreams++ achieves significantly higher success rates and more human-like driving behavior than traditional search-based methods.
Introduction
Autonomous driving faces a persistent challenge: how to handle the uncertainty of dynamic traffic while maintaining safety and interpretability. Classical rule-based systems are often too rigid for complex interactions, while end-to-end deep learning models often lack transparency. The FFStreams++ framework addresses this by integrating symbolic logic with probabilistic prediction, focusing on maneuvers where predicting other agents' intents is critical for safety and efficiency.
The FFStreams++ Architecture
The core of the system is a multi-stage pipeline that translates environmental data into actionable, optimized plans.
- Trajectory Prediction (QCNet): The system uses a Query-Centric Network (QCNet) pre-trained on the Argoverse 2 dataset. It predicts the six most likely future trajectories for surrounding obstacles, providing the planner with a probabilistic "map" of future risks.
- Symbolic Reasoning (PDDL 2.1): High-level "rules of the road" are encoded using the Planning Domain Definition Language. This allows the vehicle to reason about discrete states (e.g., "yielding" vs. "proceeding") before calculating specific paths.
- Maneuver-Specific Streams: These generate candidate trajectories in Frenet coordinates, optimizing for passenger comfort by minimizing jerk (the rate of change of acceleration) and travel time.
- Fast-Forward (FF) Planner: A heuristic search algorithm finds the optimal sequence of maneuvers by evaluating a relaxed planning graph, ensuring efficient decision-making in real-time.
Experimental Evaluation
The framework was tested using the CommonRoad benchmark suite, specifically focusing on a complex intersection in Nuremberg (DEU_Nuremberg-39) and a highway scenario on US-101 (USA_US101-22). These scenarios represent high-stakes environments requiring precise timing and intent estimation.
Key Results
FFStreams++ significantly outperformed a baseline search-based planner utilizing 2,697 motion primitives.
As shown in the table, the success rate for unprotected left turns improved from 53% to 84%. Beyond raw success rates, the Occupant’s Preference Metric (OPM) revealed that FFStreams++ produced behavior consistent with a "Normal Driver," whereas the baseline often exhibited "Aggressive" tendencies or planned inefficient stops. In qualitative tests, FFStreams++ successfully yielded when necessary and accelerated smoothly into gaps, mimicking human decision-making.
Conclusion and Future Work
FFStreams++ demonstrates the power of hybrid planning. By leveraging neural networks for prediction and symbolic logic for decision-making, it achieves a balance of safety, efficiency, and human-like comfort. However, the system currently requires high-fidelity HD maps. Future work will focus on reducing this dependency and optimizing the computational overhead for real-time deployment on standard automotive hardware.
Summary updated