Safe Robot Navigation in Indoor Healthcare Workspaces

Healthcare workspaces would greatly benefit from the employment of robotic assistants in both clinical and non-clinical tasks. However, despite their advantages, a major shortcoming for the deployment of robots limiting their widespread acceptance by the market is the fact that existing robotic solutions were originally designed for large industrial and warehouse spaces. These are characterized by structured spaces and predictable environments, where robots move along predefined paths and interaction with humans is typically not required. Herein, we examine state-of-the-art computer vision methods that enable robots to detect the presence and identify the type of dynamic obstacles inside their visual field and adapt their navigation accordingly. To achieve this goal, we trained our robots using contemporary deep learning methods (namely YOLO-You Only Look Once architecture and its variations) and obtained promising results in both human and robot detection. For that purpose, a newly constructed dataset consisting of robot images was used, complementing the well-known COCO dataset. Overall, the present study contributes towards the key objective of safe robot navigation in healthcare spaces and underpins the wider application of studies on Human-Robot Interaction in less structured environments.