Towards generic object property estimation using Unsupervised Renforcement Learning

Overview & Synopsis

Scene understanding for an autonomous agent is essential to perform tasks in an environment often made up of a multitude of objects of complex appearances and dynamics. In this paper, we propose a generic method to train an agent to physically interact with an object to extract several of its physical properties. In a real case, the agent may face unidentified objects with no prior knowledge. For this purpose, we used an algorithm from Unsupervised Reinforcement Learning (URL) to learn such an interaction strategy. The observations acquired by the agent along its trajectory are used to predict several properties with a model learned in a supervised manner. We applied our method in a simulated environment with a mobile robot only capable of interacting through pushing in order to extract dynamic parameters such as mass, shape, capacity to roll and slide. The results show that URL is able to provide diversity in trajectories to extract properties even in noisy conditions.