Enhancing Path Planning Performance through Image Representation Learning of High-Dimensional Configuration Spaces

TL;DR

The paper addresses accelerating path planning in high-dimensional configuration spaces with unknown obstacles by learning a biased sampling distribution for waypoints. It introduces a WGAN-GP conditioned via a forward diffusion latent and an RGB-D+-based image representation to model the distribution over C_free, supplemented by affinity propagation to compress waypoint structure into exemplars for single-query waypoint generation. The planner integrates this biased sampler with RRT/RRT* while including a failure-detection mechanism to revert to uniform sampling, thereby preserving probabilistic completeness, and demonstrates improved planning time and success rate under tight time constraints on Baxter 7-DOF tasks. Results show substantial gains in time-constrained scenarios, with code and datasets available for reproduction, underscoring the practical impact for real-time robotics.

Abstract

This paper presents a novel method for accelerating path-planning tasks in unknown scenes with obstacles by utilizing Wasserstein Generative Adversarial Networks (WGANs) with Gradient Penalty (GP) to approximate the distribution of waypoints for a collision-free path using the Rapidly-exploring Random Tree algorithm. Our approach involves conditioning the WGAN-GP with a forward diffusion process in a continuous latent space to handle multimodal datasets effectively. We also propose encoding the waypoints of a collision-free path as a matrix, where the multidimensional ordering of the waypoints is naturally preserved. This method not only improves model learning but also enhances training convergence. Furthermore, we propose a method to assess whether the trained model fails to accurately capture the true waypoints. In such cases, we revert to uniform sampling to ensure the algorithm's probabilistic completeness; a process that traditionally involves manually determining an optimal ratio for each scenario in other machine learning-based methods. Our experiments demonstrate promising results in accelerating path-planning tasks under critical time constraints. The source code is openly available at https://bitbucket.org/joro3001/imagewgangpplanning/src/master/.

Figures (5)

• Figure 1: Proposed architecture to learn the waypoints of a collision-free path from the robot's WS.
• Figure 2: Projection of the first two joints of the Baxter manipulator's arm and its affinity points $\bar{\bm{q}}$. The circle represents the standard deviation $\sigma$. A fixed $\sigma$ could miss the original points around the exemplars, as shown in Fig. \ref{['fig:originalCF']}. However, if we increase $\sigma$, we eventually also increase the probability of sampling the missing $\mathcal{C}_{free}$-configurations, as shown in Fig. \ref{['fig:originalCF1']}
• Figure 3: In the case of a constrained path, such as the shortest path, there is already an inherent ordering relationship between the configurations. In this example, we observe an almost continuous gradient between the different states in the matrix representation of a collision-free path, each entry of the waypoints are scaled.
• Figure 4: Conventional and proposed datasets for waypoints in $\mathcal{C}_{free}$ generative model training. In Fig. \ref{['fig:figureDataseta']}, we show how the dataset during training would repeat the same working space RGB-D for every waypoint sampled from $\mathcal{C}_{free}$; thus requiring to increase the number of samples to be processed in the same epoch. In Fig. \ref{['fig:figureDatasetb']}, the representation only requires a paired matrix of centroids to represent all the waypoints, which decreases the total number of samples per epoch and decreases the time of learning.
• Figure 5: Extra channel to add conditioning to the generator. For instance, when training the model to generate waypoints for the shortest path, we include the start and end states of the path in this extra channel; the condition consists of 14 entries, where 7 corresponding to the start state and 7 to the end state. Each row represents different encoding.