Identifying Objects’ Inertial Parameters with Robotic Manipulation to Create Simulation-Ready Assets

Abstract

Real2Sim is the problem of simulating objects and scenes via real world data, allowing a robot to imagine future interactions with its environment. However, many existing approaches either do not consider the dynamics of objects being simulated or make assumptions about their mass distributions. In this work, we aim to make use of robotic arm payload identification techniques in order to enhance the dynamic accuracy of objects generated from a Real2Sim pipeline for manipulation tasks. While the payload identification literature is vast, applying these methods in practice has various challenges and limitations. Upon implementing these techniques, we gain understanding of best practices in the engineering sense. We hope that these methods can be used to provide ground truth data for other robot learning tasks on the road towards generalized dynamic intuition.