DEEP LEARNING

DeformableRavens — New Google AI Toolbox for Robotics

Recently Google AI team presented a set of neural networks, allowing robots to be trained to move one-, two- and three-dimensional deformable objects. The code opens up opportunities for increasing the level of automation in manufacturing.

An example of a trained Transporter Network policy in action on the bag-items-1 task. Source: arxiv

About the problem

The difficulty that arises when manipulating a deformable object lies in the impossibility of completely setting its configuration. For example, to describe the location of a rigid cube in three-dimensional space, it is sufficient to indicate the position of a fixed point relative to its center, but in an object such as a cloth, the position of all its points changes relative to each other. Even in the presence of an accurate description of the state of a deformable object, the problem of reconstructing its dynamics remains. This makes it difficult to predict the future state of a deformable object after an action has already been applied to it, which is often important for multi-stage planning algorithms.

This article presents the open-source DeformableRavens benchmark, which includes 12 tasks related to manipulating one-dimensional (cables), two-dimensional (fabrics), and three-dimensional (bags) objects. Each comes with a scripted expert demonstrator that succeeds with high probability…