Online Acquisition of Simulatable Rigid Object Models

Abstract

How can we build a robot that operates autonomously in a home environment over long periods of time? A key requirement is the ability to perceive and understand its surroundings, including the objects it will interact with. This thesis investigates how a robot can reconstruct previously unknown objects and integrate them into a physics simulation for planning. We explore two methods for reconstructing the 3D geometry of objects and test their performance in simulation and in real-world experiments. Our results demonstrate that a learned depth model enables 3D reconstruction of unknown objects and their successful integration into simulation environments. Additionally, we investigate methods for estimating an object’s inertial parameters, using its reconstructed mesh and through manipulation.