From Component to System: A Task-Unified Planning System with Planning-Oriented Predictor

Autonomous driving is developing rapidly and has become a hot topic in both industry and research. The planning system plays a crucial role in meeting the requirements of autonomous driving. However, current planning system designs may not effectively serve planning tasks. A typical modular planning system offers high interpretability and flexibility. However, it may cause task-agnostic problems between the upstream predictor and the downstream planner. End-to-end driving systems have a natural advantage in achieving system-wide integration, but their poor interpretability poses safety risks. Therefore, in this paper, a task-unified planning framework is proposed to inspire the current prediction-planning paradigm. In this architecture, driving tasks are first modeled. Then, the predictor and planner are jointly designed and optimized based on these tasks. Finally, during the actual planning process, the upstream and downstream components remain relatively independent to allow for flexible adjustments. The core of this architecture is a planning-oriented predictor named POP, which fully retains the advantages of modular systems by optimizing the predictor to meet driving requirements. Comprehensive experiments demonstrate its effectiveness. Compared to typical modular systems, POP-based framework shows significant improvements in planning tasks, particularly in collision avoidance, ensuring system safety without compromising driving efficiency or comfort.