Towards Image-based Material Property Estimation Using The Robot Quadruped Material Database (RQMD)

This research paper presents an image-based approach for material property estimation that is used with a quadruped robot, to estimate the material property of the terrains it will be encountering. We introduce a comprehensive dataset, the Robot Quadruped Material Database (RQMD), comprising a diverse range of material samples. The dataset consists of RGB images captured with a Raspberry Pi camera, depicting various terrains such as asphalt, brick, grass, gravel, and tiles. Moreover, we introduce a novel method for extracting friction forces between the given terrains, and the leg of the robot. Incorporating the friction force introduces an additional aspect to the quadruped's decision-making process. Just as humans benefit from this extra information, it can aid in making decisions related to locomotion, maneuverability, and the quadruped's interaction with its surroundings. Additionally, we propose a predictive model that utilizes deep learning techniques for high-accuracy (95%) material property estimation. The model takes image inputs and employs a convolutional neural network (CNN) architecture to learn meaningful representations of material surfaces. Through extensive training and validation, our model achieves remarkable accuracy in predicting material classes. The contributions of this study have significant implications for various fields, such as robotics, material science, and industrial applications.