Focus on Impact: Indoor Exploration With Intrinsic Motivation

Exploration of indoor environments has recently experienced a significant interest, also thanks to the introduction deep neural agents built in hierarchical fashion and trained with Deep Reinforcement Learning (DRL) on simulated environments. Current state-of-the-art methods employ dense extrinsic reward that requires complete priori knowledge layout training environment learn an effective exploration policy. However, such information is expensive gather terms time resources. In this work, we propose train model purely intrinsic signal guide exploration, which based impact robot’s actions its internal representation environment. So far, impact-based rewards have been employed for simple tasks procedurally generated synthetic countable states. Since number states observable by agent realistic non-countable, include neural-based density replace traditional count-based regularization estimated pseudo-count previously visited The proposed approach outperforms DRL-based competitors relying surpasses computed layouts. We show robot equipped seamlessly adapts point-goal navigation real-world deployment.