Generalizable Data-driven Models for Personalized Shared Control of Human-Machine Systems

Broad, Alexander Sherman

doi:https://doi.org/10.21985/n2-xab5-dj29

Work

Generalizable Data-driven Models for Personalized Shared Control of Human-Machine Systems

Public

Download PDF

Download All Files (.zip)

The theory of how humans and machines control and communicate with each other is at the core of the scientific field known as Human-Robot Interaction (HRI). Researchers in this sub-discipline of robotics are therefore particularly interested in developing methods to chuppahreduce the inherent friction in this communication and control channel. Just as can be observed in the analogous problem of collaboration between two human partners, solutions in this space require a tight coupling between a human partner and an autonomous partner. A conceptual framework that describes this exact relationship is known as shared control (SC). Shared control defines an abstract link between a set of partners (often a human operator and an autonomous agent) that are both responsible for providing control information to the same robotic device. This paradigm is especially useful as a method of extending the physical capabilities of a human operator, while simultaneously considering important constraints defined by the user and environment. This dissertation is largely motivated by applications of shared control in the fields of assistive and rehabilitation medicine. Therefore, this thesis develops shared control solutions that are designed specifically to improve, or restore, a human operator's ability to control complex mechanical devices. Example motivating systems include powered wheelchairs, exoskeletons, and robotic manipulators. In addition to increasing a human operator's capabilities, a particularly desirable attribute of any interactive system in assistive and rehabilitation medicine is the acceptance, and enjoyment, of the human-in-the-loop. For this reason, the SC algorithms described in this dissertation allocate the majority of the control authority to the human partner, while the autonomous partner is mainly responsible for providing control information to improve the stability and safety of the joint human-machine system. The specific techniques described in this dissertation are motivated by the desire to generalize solutions in shared control to generic pairs of human and machine partners, while simultaneously developing a decision making framework that is responsive to the individual human-in-the-loop. To address this desire, this thesis introduces the notion of data-driven model-based shared control (MbSC). Data-driven MbSC extends the efficacy of standard shared control systems to scenarios in which we do not have any prior knowledge of the system dynamics or the human operator. Instead, data-driven MbSC relies on techniques from (1) machine learning to gain an understanding of the joint human-machine system from observation, and (2) optimal control (OC) to develop a control policy for the autonomous partner. The shared control system then allocates authority to each partner to improve desired outcomes (e.g. task-success, stability, and/or safety). Additionally, this dissertation describes data-driven techniques that further personalize the interaction paradigm to the individual human-in-the-loop. The proposed methodology uses a representation of the autonomy's trust in the human partner's control skill learned from observation data. This data-driven metric is then used to modulate the control authority granted to each partner in real-time. Taken together, the techniques described in this thesis describe a generalizable solution to the shared control problem that can be personalized to the individual human-in-the-loop to improve the capabilities of the joint system.

Creator