Learning robot arm trajectories using deep learning for cleaning tasks

Abstract

Today the robots have already started to assist humans in the daily tasks like vacuum cleaning and lawn mowing. The time is not too far when humanoid robots would work together with humans in their daily lives. Service robotics has already started to flourish with robots serving food in the restaurants and interacting with people on airports and shopping centers. With robots getting better and better with the hardware, some big challenges still re- main on the software end. There is a need of reliable and robust cognitive approaches that can help the robot to perceive the environment, reason and plan actions dynamically. This work focuses on improving the cognitive abil- ities of the humanoid robot Rollin’ Justin and help the robot to clean its environment. To enable the robot to perceive its environment, first a dirt de- tection algorithm is developed using deep learning that helps in classification and localization of dust particles using camera images. To clean the dirty surface with motions like wiping and scrubbing, a new and unique approach using learning from demonstration is developed which enables the robot to learn new trajectory shapes by learning joint angles and Cartesian position of end effectors using machine learning and deep neural networks. As the trajectories are learned from human demonstrations, the same algorithm can be used to learn different types of trajectories. The resulting wiping motions learned are smoother along the surface to be wiped than the ones generated by inverse kinematics solver following a planned path and thus enabling the robot to clean effectively. Later a new approach is proposed to combine the predicted trajectories of two machine learning algorithms that can be used in future to improve the learned trajectory with the help of feedback from sensors.