MIMP: Modular and Interpretable Motion Planning Framework for Safe Autonomous Driving in Complex Real-world Scenarios

Motion planning for autonomous vehicles in complex, real-world urban scenarios is a fundamental challenge in autonomous driving. To this end, we present MIMP, a Modular and Interpretable Motion Planning framework tailored for operation in such complex scenarios. Our approach consists of three key modules: trajectory generation, trajectory scoring with a trainable cost volume, and a safety filter. In the trajectory generation module, a wide range of driving behaviors is covered by generating a large set of potentially viable trajectories for the ego vehicle. To score these trajectories, we use a deep learning model, which learns a spatio-temporal cost volume to assess all trajectories in real-time. Finally, a safety filter module ensures safety in a deterministic and verifiable manner by checking for compliance with the drivable area and the absence of collisions with other agents in their future positions, obtained with a simple projection module. The trajectory with the lowest cost that passes the safety filter is the final plan, without any additional adjustments. Our results in closed-loop testing closely match those of other top-performing methods on the nuPlan benchmark and outperform them in most challenging scenarios. We emphasize the simplicity of our three building blocks, demonstrating the potential of an elegant and straightforward approach for motion planning.