Cognitive Intelligence for Professional Cleaning Robots

Abstract:
This paper presents a holistic scene perception approach that enhances the cognitive intelligence of professional cleaning robots to understand complex and dynamic environments. It builds upon a recently published CNN to detect and distinguish between dirt and objects on the floor. A multi-frame fusion of the detection results helps to increase the detection accuracy and provides semantic scene knowledge. Moreover, it enables the robot to perceive and avoid even small objects on the floor being barely detectable in the spatial sensor data. Lastly, the approach generates long-term dirt maps from scene knowledge to adjust the coverage strategy, e.g., to clean dirty areas more frequently. The integration of the holistic scene perception with state-of-the-art techniques for navigation, such as SLAM and coverage planning, and the deployment into a demonstrator, exemplifies the improvements on professional cleaning tasks.