Active Auditory Exploration for Identifying Object Contents

Intelligent robots need to acquire a target object from the same type of container filled with different kinds of objects. The actions selected by robots are critical to quickly and accurately recognizing the type of objects inside the container. In this paper, we propose an active auditory exploration method based on reinforcement learning, which enables the robot to actively explore the operational behavior of interest and establish the coupling relationship between perception and action to reduce the ambiguity of target recognition. The robot interacts with visually indistinguishable bottles by adopting multiple action behaviors to generate sound data from which the perceptual model learns to classify object contents. This is similar to human beings actively exploring and accumulating experience through sound in the environment where vision cannot be judged. In addition, we also use passive strategies to compare and analyze the impact of inhibition behavior on object recognition. Experimental results show that our method successfully learns effective class recognition strategies and actively selects actions to further improve recognition efficiency.