By IEEE Spectrum
June 29, 2017
A dataset of 6.7 million robust point clouds and grasps can train your neural network to reliably pick up objects.
Credit: AUTOLAB/UC Berkeley
University of California, Berkeley professor Ken Goldberg on Wednesday announced the release of a massive dataset for Dex-Net 2.0, a project employing neural networks to develop reliable robot grasping across an array of rigid objects.
The dataset houses 6.7-million point object point clouds, along with parallel-jaw gripper poses and a robustness estimate of how likely it is the grasp will successfully lift and carry the object.
Working with AUTOLAB’s Jeff Mahler, Goldberg trained a convolutional neural network to predict the robustness of a specific grasp on a given object, using “a probabilistic model to generate synthetic point clouds, grasps, and grasp robustness labels from datasets of [three-dimensional] object meshes using physics-based models of grasping, image rendering, and camera noise.”
Experimenting with one robot yielded 93% reliable grasp planning, and Goldberg says the dataset “can be a useful resource to train deep-neural networks for robot grasp planning across multiple different robots.”
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA
No entries found
from Communications of the ACM: Artificial Intelligence http://ift.tt/2ttIHyA