Adaptive Trajectory Learning With Obstacle Awareness For Motion Planning

In motion planning, efficiently navigating from a start state to a goal state in spaces with narrow passages remains a significant challenge. Recently, learning-based methods have attracted considerable attention owing to their higher inference speeds compared to traditional approaches. However, the variability in state distribution on the expert path hinders the training of neural networks, while the overly dense states may lead to redundant decision iterations and unsatisfactory planning efficiency. In this paper, we present a novel deep learning framework for motion planning, termed Adaptive Trajectory Learning with Obstacle Awareness (ATOA). Instead of performing the conventional state-wise supervision that approaches the next state, we propose to learn the trajectory along the expert path. This mechanism not only mitigates the model’s dependence on the expert paths but also has the potential to yield more effective planning solutions. Additionally, obstacle information is explicitly integrated by penalizing predictions with obstacle collisions. To further enhance the planning success rate, we introduce a confidence-driven path correction (CDPC) module to adjust the infeasible local paths. Extensive experiments demonstrate the effectiveness and superiority of ATOA compared to prior approaches in handling complex scenarios. The code will be released to promote further studies.