A Differentiable Constraint-Aware Motion Planning Module for Unsupervised Trajectory Optimization

Engineering notes

Key topics: autonomous driving, end-to-end, motion planning, trajectory planning, nuscenes, planning, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous driving requires planning trajectories that are not only goal-oriented but also compliant with safety and feasibility constraints such as lane boundaries, inter-agent distances, and motion smoothness. However, in dynamic and uncertain environments, obtaining ground-truth trajectories for supervised learning can be costly or impractical. In this paper, we propose a constraint-aware, supervision-free trajectory planning framework that could learn safe and executable behaviors directly from violation-prone data. Our approach outputs control actions that are rolled out using a kinematic model to generate trajectories, which are then optimized via a differentiable loss incorporating learnable Lagrangian penalties. In particular, we develop a lane boundary constraint based on signed lateral deviations and enforce it through a smooth penalty formulation. Experiments on the nuScenes Open Motion Dataset demonstrate that our method ensures strict compliance with road boundaries while staying close to reference goals, and it can be seamlessly integrated as a plug-in module with End-to-End pipelines.

Links and sources

• Official / arXiv page

Comments