Proceedings of the Annual Meeting of the Cognitive Science Society

When encountering unfamiliar physical objects, children andadults often perform structured interrogatory actions such asgrasping and prodding, so revealing latent physical propertiessuch as masses and textures. However, the processes drivingand supporting these curious behaviors are still largely mys-terious. In this paper, we develop and train an agent able toactively uncover latent physical properties such as the massand force of objects in a simulated physical “micro-world’.Concretely, we used a simulation-based-inference frameworkto quantify the physical information produced by observationand interaction with the evolving dynamic environment. Weused model-free reinforcement learning algorithm to train anagent to implement general strategies for revealing latent phys-ical properties. We compare the behaviors of this agent to thehuman behaviors observed in a similar task.