This paper presents Ego-Foresight, a novel method inspired by human movement prediction, to address the sample efficiency problem of deep reinforcement learning (RL). To overcome the large training data requirements of conventional RL, we take an approach that separates the agent and its environment from each other. However, unlike previous studies, we learn the agent-environment interaction using the agent's movements themselves, without any supervised signals. Ego-Foresight enhances the agent's perception ability through self-supervised learning via visual-motor predictions, enabling it to predict agent movements from simulated and real-world robot data. By integrating it with model-free RL algorithms, we demonstrate improved sample efficiency and performance.
