Hierarchical Landmark Policy Optimization for Visual Indoor Navigation

In this paper, we study the problem of visual indoor navigation to an object that is defined by its semantic category. Recent works have shown significant achievements in end-to-end reinforcement learning approach and modular systems. However, both approaches need a big step forward be robust practically applicable. To solve insufficient exploration scenes make more semantically meaningful, extend standard task formulation give agent easily accessible landmarks form room locations those types. The availability allows build hierarchical policy structure achieve success rate 63% on validation photo-realistic Habitat simulator. hierarchy, low level consists separately trained RL skills high deterministic policy, which decides skill needed at moment. Also, show possibility transferring real robot. After bit training reconstructed scene, robot shows up 79% SPL when solving navigating arbitrary object.