Transformer Based Approach for Sample Generation in Motion Planning

Published In

2023 IEEE 19th International Conference on Automation Science and Engineering (CASE)

Document Type

Citation

Publication Date

8-26-2023

Abstract

Efficient motion planning algorithms are critical in real-world robot applications, but existing methods become less effective as the robot's dimensionality and workspace increase, primarily due to the computational cost of collision checking during sample generation. We propose a framework that generates robot samples directly for the planning problem to overcome the high cost of expensive operations in sampling-based motion planning. The framework utilizes a combination of an AutoEncoder (AE) and a Transformer-based model, which efficiently generates valid robot samples using the output from the AE, capturing a latent space representation of the robot's environments represented as point clouds. The framework is tested on various ground and air robot planning problems in 3D environments and performs better than comparable methods across all the tested problems. It also demonstrates the ability to generalize to new and unseen environments.

Rights

©2023 IEEE

DOI

10.1109/CASE56687.2023.10260470

Persistent Identifier

https://archives.pdx.edu/ds/psu/40924

Publisher

IEEE

Share

COinS