Neural Informed RRT*: Learning-based Path Planning with Point Cloud State Representations under Admissible Ellipsoidal Constraints

TL;DR

Problem: accelerate convergence of asymptotically optimal path planning in cluttered environments while preserving probabilistic completeness. Approach: Neural Informed RRT* (NIRRT*) combines Informed RRT* with a point-based network that processes a free-space point cloud; Neural Focus constrains inference to the ellipsoidal focus region defined by $c_{\textrm{curr}}$, and Neural Connect enforces connectivity among guided states. Contributions: (1) PointNet++-based guidance on point clouds, (2) Neural Focus integrating learning with informed sampling, (3) Neural Connect ensuring connectivity to improve sample efficiency. Findings: across 2D/3D benchmarks and real-world TurtleBot deployment, NIRRT* variants converge faster and reach closer to the optimal cost than RRT*, IRRT*, or NRRT*-GNG. Significance: provides a scalable, sensor-friendly framework for learning-guided, admissible-sample path planning applicable to mobile robots in dynamic environments.

Abstract

Sampling-based planning algorithms like Rapidly-exploring Random Tree (RRT) are versatile in solving path planning problems. RRT* offers asymptotic optimality but requires growing the tree uniformly over the free space, which leaves room for efficiency improvement. To accelerate convergence, rule-based informed approaches sample states in an admissible ellipsoidal subset of the space determined by the current path cost. Learning-based alternatives model the topology of the free space and infer the states close to the optimal path to guide planning. We propose Neural Informed RRT* to combine the strengths from both sides. We define point cloud representations of free states. We perform Neural Focus, which constrains the point cloud within the admissible ellipsoidal subset from Informed RRT*, and feeds into PointNet++ for refined guidance state inference. In addition, we introduce Neural Connect to build connectivity of the guidance state set and further boost performance in challenging planning problems. Our method surpasses previous works in path planning benchmarks while preserving probabilistic completeness and asymptotic optimality. We deploy our method on a mobile robot and demonstrate real world navigation around static obstacles and dynamic humans. Code is available at https://github.com/tedhuang96/nirrt_star.

Figures (6)

• Figure 1: Solutions of a 2D random world found by RRT* karaman2011sampling, Informed RRT* (IRRT*) gammell2014informed, Neural RRT* with Grid-based Network Guidance (NRRT*-GNG) wang2020neural, and Neural Informed RRT* with Point-based Network Guidance (NIRRT*-PNG). NIRRT* effectively integrates IRRT* and point-based network, so IRRT* helps point-based network focus guidance state inference on the important region for solution improvement, and point-based network helps IRRT* sample critical states in the admissible ellipsoidal subset for convergence acceleration.
• Figure 2: Guidance state inference by point-based network. Red is start, yellow is goal, blue is free states, and orange is guidance states.
• Figure 3: (a-b) Neural Focus in a 2D narrow passage and (c-h) Neural Connect in a 2D random world. Green dots denote states visited by Breadth First Search (BFS). Circles of larger size around green dots denote boundary points. The colors of circles denote the heuristic scores, where brighter colors represent higher scores. The boundary point which is $x_{\textrm{start}}^{i+1}$ or $x_{\textrm{goal}}^{i+1}$ has a blue marker on the circle. The orange line denotes the path found by BFS which represents the connectivity of $X_{\textrm{guide}}$.
• Figure 4: Center block and narrow passage experiments.
• Figure 5: Experiment results. Center block: (a) The average number of iterations to find a path within 2% of the optimal cost for different map widths; (b) Comparison of the number of iterations IRRT* and NIRRT*-PNG(FC) take to find a path within 2% of the optimal cost for each center block problem; (c) The average number of iterations to find a path within the specified tolerance of the optimal cost. 3D Random world: (d) The average path cost relative to the initial solution of RRT* at different numbers of iterations after finding an initial solution. Narrow passage: (e) The average number of iterations to find a path better than flanking the obstacle for different gap heights; (f) Comparison of the number of iterations IRRT* and NIRRT*-PNG(FC) take to find a path better than flanking the obstacle for each narrow passage problem. 2D Random world: (g) The average path cost relative to the initial solution of RRT* at different numbers of iterations after finding an initial solution; (h) Comparison of the number of iterations IRRT* and NIRRT*-PNG(FC) to find an initial solution for each random world problem. Error bars denote 95% confidence interval. The error bars are not plotted for random worlds for clarity of figures. NRRT*-GNG is not implemented for center block and random world 3D due to incompatible grid sizes and incompatible number of dimensionality.