A Kernel-based Approach to Learning Contact Distributions for Robot Manipulation Tasks

Abstract

Manipulation tasks often require robots to recognize interactions between objects. For example, a robot may need to determine if it has grasped an object properly or if one object is resting on another in a stable manner. These interactions usually depend on the contacts between the objects, with different distributions of contacts affording different interactions. In this paper, we address the problem of learning to recognize interactions between objects based on contact distributions. We present a kernel-based approach for representing the estimated contact distributions. The kernel can be used for various interactions, and it allows the robot to employ a variety of kernel methods from machine learning. The approach was evaluated on blind grasping, lifting, and stacking tasks. Using 30 training samples and the proposed kernel, the robot already achieved classification accuracies of 71.9%, 85.93%, and 97.5% for the blind grasping, lifting and stacking tasks respectively. The kernel was also used to cluster interactions using spectral clustering. The clustering method successfully differentiated between different types of interactions, including placing, inserting, and pushing. The contact points were extracted using tactile sensors or 3D point cloud models of the objects. The robot could construct small towers of assorted blocks using the classifier for the stacking task.