PROQUEST: 1873484783
Robotic devices have been used in post-stroke or trauma neurorehabilitation process for years. Due to the uncertain, complex, and changing physical human-robot-interaction dynamics involved and task-specific objectives for rehabilitation, control of this type of devices remains a vast area of research. This thesis seeks systematicness in rehabilitation robot control design by developing control frameworks for upper- and low-limb rehabilitation devices that can be generally applied on multiple devices, using theoretical and experimental methodologies.
The thesis examines the control difficulties and provides solutions for two specific groups of rehabilitation robots that are widely used: the upper-limb rehabilitation robot for reaching and tracking tasks that consists of a robot arm linked to human arm in series and the powered ankle exoskeleton for human gait training. To address systematic control for these two groups of robots, the thesis investigates several separate problems.
The stability of human-robot interaction in upper-limb rehabilitation robots for reaching and tracking tasks training is solved in this thesis in a systematic way by means of nonlinear control theory. In addition to stability, the control framework developed ensures human safety and realizes different operations modes during training: humandominant mode in which human is partially paralyzed and given relative larger freedom to exercise to enhance mobility, robot-dominant mode in which human movements are less reliable and more robot assistance is needed and system-stop mode in which the actual human movements drift too far from the desired trajectories and should be stopped to avoid human injuries. This passivity-based, provably stable control framework can be extended to various upper-limb rehabilitation devices as a control base. Functionality of the controller was tested and confirmed effective through computer simulation and hardware experiments.
This thesis also develops a robust torque control structure for tethered ankle orthoses in human walking driven by series elastic actuators in face of the complicated and time-varying system dynamics. Major challenges of this control problem lie in the complicated and changing human-robot interactive dynamics, state-dependent rapidly varying torque objectives, and complex friction dynamics in Bowden cable transmission. This thesis designs, implements and tests a series of prominent torque controllers for powered lower-limb exoskeleton during human walking, including the combination and variation of model-free, model-based, feedback and feed-forward controllers. According to the experimental results, model-based control elements have little or even negative benefits on control performance due to the changing dynamics while requiring great efforts on system identification and high computation power. Continuous integral actions are not effective either due to the nonlinear and changing system dynamics. It is shown that the combination of model-free, integration-free feedback control and iterative learning is best suited for this type of devices used in cyclic applications like locomotion. The results support this controller in various robotic systems with properties similar to ankle exoskeletons such as nonlinear, complicated and changing dynamics, and cyclic operations, e.g., legged robots, lower-limb prostheses.
With suitable torque control architecture identified for ankle exoskeletons during human walking, this thesis also investigates by theory and experiments the interactions of actuator compliance, desired torque profile and control gains for the sake of further improving the real-time torque tracking performance. The theoretical analysis and large amount of experimental results show that for a desired quasi-stiffness generated by relating torque and ankle angle, the optimal actuator passive stiffness for control is the one that matches the desired one. Besides, for a desired passive actuator stiffness, the optimal iterative learning gain that benefits real-time tracking performance matches the inverse of passive stiffness. These results provide guidance for ankle exoskeleton system elasticity with a fixed control objective, and for iterative controller gain with a combination of fixed control objective and system elasticity.
With low-level torque control solved at a satisfactory extent, this thesis also develops an online optimization system for multivariate high level controllers in lower-limb exoskeletons from human metabolic measurements. This system enables fast and customized automation of the identification of assistance conditions that benefit human energy efficiency. Four searching algorithms, including both gradient-based and nongradient methods, were tested in simulation. Two of them, Nelder–Mead method, and Covariance Matrix Adaptation Evolution Strategy, were selected for hardware implementation due to their suitability for high dimension, high noise searching problems revealed by simulations. This system was implemented and tested on an ankle exoskeleton. Experiments demonstrated the effectiveness of both methods, especially CMA-ES, on the optimization of multi-parameter high level controller using real-time human respiratory measurements. Total evaluation time used to optimize multivariate high level controllers in human-robot interactions is greatly reduced while human metabolic rate is reduced significantly. The system can be easily extended to other human-robot interactive systems that aim for human metabolic benefits, and can be scaled to higher dimensional parameter optimization without heavily increasing evaluation time. This on-line optimization process is expected to be conducted at the beginning of each long experimental or training sessions for all subjects, involvement of which is expected to improve the corresponding operation efficiency.
In summary, this thesis systematically solved