Image-Based material property estimation

Abstract

In this paper, we introduce the Robot Quadruped Materials Dataset (RQMD), which is a collection of RGB color images as well as a friction coefficient associated with each image for 5 different material types taken in different lighting conditions. This dataset is tailored for computer vision research and specifically in the field of robotics applications. The selected surface material types are commonly traversed by robotic systems and offer a wide variety of different friction coefficients. The image data is collected using a Raspberry Pi camera as well as a Raspberry Pi embedded computer for future integration with the quadruped robot. The addition of the friction coefficient provides a new dimension to the decision-making of the quadruped. Similarly to humans, this additional information can be helpful in decision-making in tasks such as locomotion, maneuverability, and interaction of the quadruped with the environment.The friction coefficients were extracted using a DC motor with encoder. We also verify that existing Convolutional Neural Networks (CNN) architectures such as the MobileNetV2 and InceptionV3 among others can be trained on the RQMD for image classification and obtain accuracies greater than 90%.