SDT-Planner: Leveraging the Soft Decision Tree for Interpretable Motion Planning in Autonomous Driving

Although deep neural networks are increasingly applied to motion planning of autonomous vehicles in diverse scenarios, the inherent lack of interpretability raises safety, trust and regulatory concerns. In light of this, we propose an interpretable motion planner employing the Soft Decision Tree (SDT). The motion planning procedure is formulated as a classification task which hierarchically partitions the input scene using high-dimensional decision boundaries learned at each node from abundant training data. In order for the SDT to plan motions across unlimited scenarios with finite steps of partition, we represent driving scenes as object occupancy maps to allow for expressive representation and consequently fast convergence, and we define each leaf node of the SDT as a probabilistic motion distribution to categorize infinite future motions into a finite set of classes. The derived motion distribution, combined with encoded traffic rules, is then decoded with a Recurrent Neural Network-based decoder to produce future trajectories for the ego-vehicle. The proposed method is evaluated in both open and closed-loop settings. In closed-loop settings on CARLA Town05-short and Town05-long benchmarks, our method achieves state-of-the-art performance, with route completion rates of 99.41% and 99.53%, as well as driving scores of 94.91 and 78.13, respectively. Visualization analysis further reveals that the proposed planner effectively captures the long-tail distribution of driving scenes, with clear correspondences between certain leaf nodes and specific motion patterns. Furthermore, our method generalizes well beyond simulation and produces safe, rule-compliant trajectories in open-loop evaluations on a real-world dataset collected from urban roads in Wuhan, China.