Abstract
Haptic collaborative tasks are actions performed jointly by several partners, involving direct or indirect physical contact among them. A typical example of such tasks are collaborative manipulation tasks, where the partners apply forces on a same object to impose it a desired motion or bring it to a target location. Human beings learn naturally how to perform such tasks with other human partners, but implementing such behaviors on a robotic platform is challenging. When jointly manipulating an object, the partners no longer act independently, and must negotiate a common plan to perform the task. To avoid conflicts among the partners' intentions, the leader-follower model defines a task leader, who imposes a task plan to the other partners, while the latter act as follower and follow at best the intentions of the leader. This model has often been used in physical Human-Robot Interaction (pHRI). Because robotic systems have limited cognitive capabilities in comparison to human beings, a follower role has generally been assigned to robotic systems to cooperate with human operators. Recently, thanks to the increasing computational power embedded into the robots, more and more initiative has been given to robotic assistants. In some recent works, robots were sometimes even given the possibility to lead human operators. In the context of physical tasks, where the partners are in direct or indirect contact through an object and exchange mechanical energy, we believe that the haptic channel is a favored and fast way for the partners to exchange information about their intentions. Therefore, this thesis will focus on the kinesthetic aspects of collaborative tasks. The long-term aim of the project is to endow humanoid robots with the necessary haptic skills to perform collaborative tasks with a human operator as a partner rather than as a helper. The work presented here proposes solutions towards this direction. In its first part, our contribution is to extend the leader-follower model to continuous, time-varying role distributions among the partners in the context of haptic dyadic collaborative tasks. This model describes the behavior of each partner of the dyad as a variable weighting between the two extreme leader and follower behaviors. Our goal is to abstract the concept of role distribution from the implementation of the underlying controllers, and to describe the behavior of dyads using two independent functions that will shape the behavior of each partner in term of leadership. We exemplify the use of our model in a virtual reality scenario where a human operator manipulates an object in cooperation with a virtual robotic system. We also explore possible strategies to exploit it. The problem we adress is to define how the weighting between both behaviors can be adjusted automatically on a robotic system, depending on various criteria such as constraints of the robot or knowledge from human-human haptic interaction. Simulations and experiments conducted on a humanoid robot are presented to illustrate the proposed solutions. The results show that the extended leader-follower model can be applied to realize collaborative tasks with a human operator while avoiding self-collision. The model also encompasses the specialization phenomenon recently highlighted in human-human collaborative haptic tasks. We then propose to use a programming by demonstration method to teach collaborative skills to a robotic system. This method uses a probabilistic framework to encode the characteristics of the task and reproduce it autonomously. This framework is based on Gaussian Mixture Models and Gaussian Mixture Regression and has been successfully applied to various stand-alone tasks. We remind the main components of this framework and present its application to collaborative lifting tasks between a humanoid robot and a human operator. Our first contribution is the design of the experimental setup, based on a teleoperation system whith kinesthetic feedback which allows the human teacher to demonstrate the task while taking into account the constraints and sensor data of the robotic system. The main contribution, however, is the use of this methodology to attempt to assess the validity of our extended leader-follower model, by highlighting smooth switching behaviors on human partners during collaborative lifting tasks. The experimental data aquired during reproductions of the task is analyzed within this perspective. The second part of this thesis focuses on the control of humanoid robots in the context of pHRI. We examine several paradigms of interaction: interaction between two remote human partners through a tele-presence system, direct interaction between an autonomous humanoid robot and a human operator, and collaborative transportation tasks between a human operator and a humanoid robot. Behind these different paradigms of interaction lies one common problem: the generation of whole-body motion and gait in response to external forces that arise from the haptic interaction with a human operator. This thesis does not aim at tackling the problem of gait generation at the mechanical and control level. We will rather use state-of-the-art algorithms which do not consider external disturbances, and show to what extent they can be used to generate complex and intuitive collaborative behaviors. Our contributions in this part are thus to integrate impedance control and gait generation within an existing control architecture in a generic and flexible way in order to (i) use the resulting controller in various contexts, (ii) demonstrate how the basic principles of impedance control can be implemented on a complex platform biped humanoid robot while exploiting all the capabilities of such platforms and (iii), highlight the limitations of the passivity-based approaches often used in pHRI, and thereby justify further research in the field of pHRI. The work presented in this part has been integrated within a complex demonstrator where the robot walks in a teleoperated manner and performs autonomously a collaborative transportation task with a human operator.