Safety-Compliant Navigation: Navigation Point-Guided Planning with Primitive Trajectories

Abstract

In learning from demonstrations (LfD) for trajectory planning, end-to-end deep learning (DL) methods offer fast inference and adaptability to complex inputs. However, they are prone to cumulative errors due to limited expert time-series data, which poses challenges in safety-critical applications. To address this, we introduce bounded discontinuities in trajectory planning, with the bound adaptively determined via binary search. Two generative networks, trained in opposite directions, produce primitive trajectories. These are connected using the discontinuity-allowed multi-point RRT-connect (DAMP-RRT-connect) algorithm, which expands the trajectory while maintaining discontinuities within the bound. A sequence of navigation points directs the expansion. Experiments on aircraft landing and takeoff tasks at a non-towered airport demonstrate the robustness and efficiency of our approach. [Code]<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>