Abstract
Most robots are better than humans in terms of their sensorsand actuators, allowing them to have higher and better perception,and faster movements and superior strength, than humans. However,what they lack is the human ability to make intelligent high leveldecisions in unstructured environments. Therefore, many currentcomplex systems, like rescue robots, UAV/UUVs, and surgery assistiverobots, utilize ‘human in the loop’ systems, where the controlis shared by both the humans and the machines. This PhD explores‘embodied robots’ as a promising phenomenon to improvehuman-in-the-loop control.By embodied robots, we refer to robots attached to the humanoperator, and which are controlled by the human operator using firstperson feedbacks (predominantly visual and haptic) that make himperceive the robot to be part of his/her self.First, we worked with an available Wearable Robot Arms (WRA)device to design an admittance controller that would let humanoperators control and use it intuitively.Second, we developed an embodied teleoperation system for a7DoF serial manipulator by integrating the robot system with ahaptic feedback device, and virtual reality head mounted displaycoupled to a 360 degree camera.Third, for this system, we developed a procedure to estimatethe impedance of the human operator without the use ofelectromyography, and use it for impedance control of the robot,enabling users to impose both movements as well as impedance on therobot slave.Finally, we studied Learning from Demonstration (LfD) usingvisual and human motion data through our embodied system. Wepostulate that embodied systems can improve LfD by enablingkinesthetic teaching without lose of the operator’s point of viewand force sensation. Indeed in such systems it is possible to recordevery motion, vision and feeling the human would have, making itideal for the implementation of machine learning techniques. Todemonstrate this idea, we developed a task requiring humanparticipants to move through a maze in the presence of various forcefields. We then trained a deep learning network agent on thesebehaviors. By using a state of the art pre-trained Residual Networksand Gated Recurrent Units, the machine learning agent was able tolearn ‘human-like’ movements and generalize, and predict humanbehavior in trajectories that it was never trained on.We conclude this work with a discussion of the achieved resultsand future perspectives of research in the field of embodied controland Learning from Demonstration.