A Learning-Based Framework for Collision-Free Motion Planning

TL;DR

This work tackles the need for robust, real-time, collision-free motion planning in cluttered environments by learning a parameter predictor for a Circular Field–based planner. A depth-image–to–voxel perception pipeline feeds a PointNet++-based network that outputs a 36‑D planner-parameter vector $\vec{p}$, inferred via Bayesian optimization (HEBO) over many randomized scenes. The predicted gains are applied within a multi-agent CF planner, enabling real-time trajectory generation and improved generalization over classical planners, demonstrated in simulation and on a Franka Emika Panda robot. The approach reduces manual tuning, accelerates planning, and lays groundwork for reactive planning in more dynamic environments with future CUDA optimizations and higher-resolution voxel representations.

Abstract

This paper presents a learning-based extension to a Circular Field (CF)-based motion planner for efficient, collision-free trajectory generation in cluttered environments. The proposed approach overcomes the limitations of hand-tuned force field parameters by employing a deep neural network trained to infer optimal planner gains from a single depth image of the scene. The pipeline incorporates a CUDA-accelerated perception module, a predictive agent-based planning strategy, and a dataset generated through Bayesian optimization in simulation. The resulting framework enables real-time planning without manual parameter tuning and is validated both in simulation and on a Franka Emika Panda robot. Experimental results demonstrate successful task completion and improved generalization compared to classical planners.

Figures (10)

• Figure 1: Scheme illustrating the circular field (dark red) and CF force (green) generation for a spherical obstacle (light red) in 2D.
• Figure 2: System diagram of the proposed learning-based approach for collision-free motion planning. Randomized scenes are used to generate optimal hyperparameter labels via Bayesian optimization on robot PMAF-generated trajectory costs. These labeled pairs are used to train a neural network that predicts suitable hyperparameters for novel input scenes.
• Figure 3: PointNet Architecture. The classification network takes n points as input, applies input and feature transformations, and then aggregates point features by max pooling. The output is classification scores for k classes
• Figure 4: PointNet ++ Architecture. The adapted model does not use a segmentation layer and replaces the final classification layer with k output classes by a fully connected layer with 36 output neurons
• Figure 5: First and last frames of the simulation. These illustrate the initial and final configurations of the task. The goal position for the robot’s flange and TCP are shown in yellow and light blue, respectively. TCP trajectories over time are shown in green.