Active Object Searching on Mobile Robot Using Reinforcement Learning

摘要

With the fast development of hardware computation, visual object detection has achieved a prosperous development in both industrial application and academic researching. Beneficial from the great success in the field of artificial intelligence, mobile robot research has received much inspiration for autonomy path planning. Although state-of-art object detection algorithms, such as YOLO, Mask-RCNN, can detect and classify targets from robot view, complicated tasks require an autonomous robot that has the capability to adjust its viewpoint and self-decide anchor targets to get better detection results. In this paper, a method is proposed to help an agent learn what actions to take according to detected objects in its viewpoint. The agent is also able to search objects in different rooms and indoor environments only with prior knowledge of potential objects in a room. Experimental results indicate that our active searching approach can help agents learn active room selection and adjust viewpoint to receive better object detection results.