Combining Machine Learning and Sampling-Based Search for Multi-Goal Motion Planning with Dynamics

TL;DR

The paper tackles multi-goal motion planning in obstacle-rich environments under robot dynamics by integrating sampling-based planning, high-level reasoning via TSP tours, and machine-learning–driven cost estimation. It introduces a motion-tree framework augmented with ML-predicted FromToCost, forms groups to reduce TSP calls, and uses TSP-guided tours to prioritize low-cost expansions. A FromToCost function blends predicted single-goal distance and runtime to reflect obstacle and dynamics difficulty, enabling efficient, scalable planning. Experimental results with a vehicle model show significant runtime improvements over state-of-the-art baselines and favorable scalability, with only modest increases in trajectory length, validating the practical impact of the approach.

Abstract

This paper considers multi-goal motion planning in unstructured, obstacle-rich environments where a robot is required to reach multiple regions while avoiding collisions. The planned motions must also satisfy the differential constraints imposed by the robot dynamics. To find solutions efficiently, this paper leverages machine learning, Traveling Salesman Problem (TSP), and sampling-based motion planning. The approach expands a motion tree by adding collision-free and dynamically-feasible trajectories as branches. A TSP solver is used to compute a tour for each node to determine the order in which to reach the remaining goals by utilizing a cost matrix. An important aspect of the approach is that it leverages machine learning to construct the cost matrix by combining runtime and distance predictions to single-goal motion-planning problems. During the motion-tree expansion, priority is given to nodes associated with low-cost tours. Experiments with a vehicle model operating in obstacle-rich environments demonstrate the computational efficiency and scalability of the approach.

Figures (7)

• Figure 1: A multi-goal motion-planning problem, where the robot has to reach each of the goals $G_0, \ldots, G_{19}$ while avoiding collisions with the obstacles. Videos of solutions obtained by our approach can be found as part of the supplementary material or by visiting the anonymous link at https://tinyurl.com/bdtsvbeu
• Figure 2: The other scenes and the vehicle model used in the experiments (scene 1 shown in Fig. \ref{['fig:Main']}).
• Figure 3: Runtime and solution distance distributions for single-goal motion-planning problems.
• Figure 4: Prediction accuracy of (a) XGBoost, (b) LightGBM, and (c) ANN on the single-goal motion-planning instances. The accuracy is computed as the percentage of test instances whose predicted value falls within the indicated tolerance (80%-20% split of the data for training and testing).
• Figure 5: Runtime results when varying the number of goals: (a) our approach, MultiGoalMP-ML (Alg. \ref{['alg:Main']}); (b) MultiGoalMP-ML[DROMOS-SingleGoal], which refers to our approach when using the single-goal version of DROMOS to generate the training datasets for the machine-learning components of our approach; (c) DROMOS; (d) MultiGoalMP-RM; (e) MultiGoalMP-ED; and (f) DROMOS-Random. The runtime includes everything from preprocessing to reporting that a solution is found. A time limit of $30$s was imposed for each run. Entries marked with $X$ indicate failure (reported runtime reached the limit).