PPNet: A Two-Stage Neural Network for End-to-end Path Planning

TL;DR

The paper tackles the challenge of end-to-end near-optimal path planning under tight time budgets. It introduces PPNet, a two-stage neural network that first segments the path space (SpaceSegNet) and then generates waypoints (WaypointGenNet), guided by the efficient data generator EDaGe-PP which provides continuous-curvature paths with analytical expressions. Empirical results show substantial gains: EDaGe-PP accelerates data generation by about 33x and increases PPNet's success rate by roughly 2x compared to traditional data-generation methods, with PPNet producing near-optimal paths in about 15.3 ms. The approach outperforms both learning-based planners and sampling-based planners in 2D scenarios, offering a practical, real-time solution for autonomous navigation and robotic motion planning.

Abstract

The classical path planners, such as sampling-based path planners, can provide probabilistic completeness guarantees in the sense that the probability that the planner fails to return a solution if one exists, decays to zero as the number of samples approaches infinity. However, finding a near-optimal feasible solution in a given period is challenging in many applications such as the autonomous vehicle. To achieve an end-to-end near-optimal path planner, we first divide the path planning problem into two subproblems, which are path space segmentation and waypoints generation in the given path's space. We further propose a two-stage neural network named Path Planning Network (PPNet) each stage solves one of the subproblems abovementioned. Moreover, we propose a novel efficient data generation method for path planning named EDaGe-PP. EDaGe-PP can generate data with continuous-curvature paths with analytical expression while satisfying the clearance requirement. The results show the total computation time of generating random 2D path planning data is less than 1/33 and the success rate of PPNet trained by the dataset that is generated by EDaGe-PP is about 2 times compared to other methods. We validate PPNet against state-of-the-art path planning methods. The results show that PPNet can find a near-optimal solution in 15.3ms, which is much shorter than the state-of-the-art path planners.

Figures (6)

• Figure 1: PPNet can find a near-optimal solution in an end-to-end way. The representative classical planner— RRT* can find a similarly optimal solution of Euclidean cost within the 5% range of the cost of the PPNet solution after the tree is expanded to have over 20,000 vertices. (a) PPNet, $t=0.015 \mathrm{s}$, $c=31.49$. (b) RRT*, $t=15.002 \mathrm{s}$, $c=30.69$. (c) RRT*, $t=30.001 \mathrm{s}$, $c=30.41$. (d) RRT*, $t=60.004 \mathrm{s}$, $c=30.48$.
• Figure 2: Four examples of the data generated by EDaGe-PP with different clearance requirements. Each figure indicates the process of generating data, which can be divided into random path generation (blue), drawing the space of the path (orange), calculating the settings of the obstacles (green), and random obstacles in the state space (black).
• Figure 3: Model overview. PPNet consists of SpaceSegNet and WaypointGenNet. SpaceSegNet can solve the first subproblem, which is space segmentation. WaypointGenNet can solve the second subproblem, which is waypoints generation.
• Figure 4: Solutions found by PPNet under different clearance settings. (a), (b), (c), and (d) clearance=1. (e), (f), (g), and (h) clearance=3.
• Figure 5: An example of PPNet, RRT*, IRRT*, BIT*, and ABIT* run on a random $\mathbb{R}^{2}$ world. Each algorithm was run until it found an equivalent solution to PPNet ($c=31.49$). (a) $t=0.015\mathrm{s}$. (b) $t=2.452\mathrm{s}$. (c) $t=0.580\mathrm{s}$. (d) $t=60.001\mathrm{s}$. (e) $t=9.564\mathrm{s}$.