Service Robot Active Object Detection based on Spatial Exploration using Deep Recurrent Q-learning Network

Abstract

Active object detection (AOD) allows robot to approach target object smoothly, but previous studies have assumed that the robot can always see the object from the start. This study introduces a new AOD model that can learn to spatial explore from any position to complete AOD task using a deep recurrent Q-learning network (DRQN). The model combines historical with current state information to accurately predict the action that should be performed at the current moment. This study also introduces a new reward function and termination condition for AOD task to ensure task completion rates and efficiency. This approach has been validated on an active vision dataset (AVD), demonstrating that service robot can actively complete AOD task from random positions through spatial exploration. Our approach has also been found to be superior to other methods.