Enhancing Service Robotics for Cleaning Tasks: A Deep Learning Approach Integrating Learning from Demonstration and Cross-Robot Knowledge Transfer

Autonomous service robots are increasingly vital in robotics research, especially as they are deployed to assist the growing elderly population. Unlike industrial robots in controlled environments, service robots must adapt to task variability and sensory feedback in unpredictable settings. This paper introduces an approach to enhance robotic cleaning by integrating Learning from Demonstration (LfD) with deep learning-based knowledge transfer. A robot learns two distinct cleaning tasks via human demonstrations, and a convolutional neural network (CNN) generalizes these demonstrations to new, unseen dirt patterns using visual feedback, enabling adaptable cleaning movements. Data augmentation and image transformations improve robustness to posture and lighting changes. Additionally, the approach allows knowledge transfer between robots in different environments. For example, a network trained on the iCub robot in Lisbon was successfully applied to the DoRo robot in Peccioli, Italy. The system autonomously performed tasks like wiping and sweeping across various environments, offering a scalable, robust framework for autonomous cleaning with applications in both household and industrial settings.