MAPLE: A Modularized Framework for Learning-Based Planners to Integrate Prediction

In autonomous driving systems, the planning module is crucial for planning feasible driving routes while predicting the movements of surrounding agents. Traditional modular planning algorithms separate these tasks, allowing for independent advancements, but often leading to error accumulation. While many approaches have sought to address this by employing end-to-end learning models or enhancing traditional planners with comprehensive prediction results, there is limited work on integrating advanced predictors with learning-based planners. To bridge this gap, we propose MAPLE, a novel modular framework that seamlessly integrates any learning-based planner with any predictor without altering the original planners. MAPLE achieves this by encoding diverse prediction formats into a planner-compatible representation, allowing learning-based planners to utilize comprehensive information about the movements of surrounding agents more effectively. This integration improves the overall planning performance, offering a robust solution to improve autonomous driving systems. Extensive experiments integrating two state-of-the-art predictors and three representative planners on the nuPlan dataset demonstrate the effectiveness and versatility of our framework, consistently improving planning scores by 3∼9%. Ablation studies and qualitative analysis further validate the design and enhanced planning capabilities of learning-based planners integrated with prediction. © 2025 IEEE.