Learning sparse and meaningful representations through embodiment

How do humans acquire a meaningful understanding of the world with little to no supervision or semantic labels provided by the environment? Here we investigate embodiment with a closed loop between action and perception as one key component in this process. We take a close look at the representations learned by a deep reinforcement learning agent that is trained with high-dimensional visual observations collected in a 3D environment with very sparse rewards. We show that this agent learns stable representations of meaningful concepts such as doors without receiving any semantic labels. Our results show that the agent learns to represent the action relevant information, extracted from a simulated camera stream, in a wide variety of sparse activation patterns. The quality of the representations learned shows the strength of embodied learning and its advantages over fully supervised approaches.