Robot Learning from Human Demonstrations for Human-Robot Synergy
Kyrarini, Maria
Universität Bremen: Physik/Elektrotechnik
robot learning from human demonstrations, reinforcement learning, human-robot interaction, human-robot synergy, assistive robots, human-robot cooperation, industrial robots
Human-robot synergy enables new developments in industrial and assistive robotics research. In recent years, collaborative robots can work together with humans to perform a task, while sharing the same workplace. However, the teachability of robots is a crucial factor, in order to establish the role of robots as human teammates. Robots require certain abilities, such as easily learning diversified tasks and adapting to unpredicted events. The most feasible method, which currently utilizes human teammate to teach robots how to perform a task, is the Robot Learning from Demonstrations (RLfD). The goal of this method is to allow non-expert users to a programa a robot by simply guiding the robot through a task. The focus of this thesis is on the development of a novel framework for Robot Learning from Demonstrations that enhances the robotsa abilities to learn and perform the sequences of actions for object manipulation tasks (high-level learning) and, simultaneously, learn and adapt the necessary trajectories for object manipulation (low-level learning). A method that automatically segments demonstrated tasks into sequences of actions is developed in this thesis. Subsequently, the generated sequences of actions are employed by a Reinforcement Learning (RL) from human demonstration approach to enable high-level robot learning. The low-level robot learning consists of a novel method that selects similar demonstrations (in case of multiple demonstrations of a task) and the Gaussian Mixture Model (GMM) method. The developed robot learning framework allows learning from single and multiple demonstrations. As soon as the robot has the knowledge of a demonstrated task, it can perform the task in cooperation with the human. However, the need for adaptation of the learned knowledge may arise during the human-robot synergy. Firstly, Interactive Reinforcement Learning (IRL) is employed as a decision support method to predict the sequence of actions in real-time, to keep the human in the loop and to enable learning the usera s preferences. Subsequently, a novel method that modifies the learned Gaussian Mixture Model (m-GMM) is developed in this thesis. This method allows the robot to cope with changes in the environment, such as objects placed in a different from the demonstrated pose or obstacles, which may be introduced by the human teammate. The modified Gaussian Mixture Model is further used by the Gaussian Mixture Regression (GMR) to generate a trajectory, which can efficiently control the robot. The developed framework for Robot Learning from Demonstrations was evaluated in two different robotic platforms: a dual-arm industrial robot and an assistive robotic manipulator. For both robotic platforms, small studies were performed for industrial and assistive manipulation tasks, respectively. Several Human-Robot Interaction (HRI) methods, such as kinesthetic teaching, gamepad or a hands-freea via head gestures, were used to provide the robot demonstrations. The a hands-freea HRI enables individuals with severe motor impairments to provide a demonstration of an assistive task. The experimental results demonstrate the potential of the developed robot learning framework to enable continuous humana robot synergy in industrial and assistive applications.
Robot Learning from Human Demonstrations for Human-Robot Synergy