Comprehensive Control Strategies for a Seven Degree of Freedom Upper Limb Exoskeleton Targeting Stroke Rehabilitation

In this dissertation control algorithms are developed and tested for the EXO-UL7 towards a control strategy for stroke rehabilitation. EXO-UL7 is a seven degree of freedom (Dof) powered upper limb exoskeleton that was initially designed at the University of Washington and further refined at the University of California Santa Cruz. Admittance control, swivel angle prediction and neural control of the device have been implemented and subjects tested performance of the device and control strategy. After an initial summery of the state of the art and details of the existing system. Each control strategy and performance from testing is explored.;Admittance control uses force sensors on EXO-UL7 to control the movement of the device by moving in the direction that user pressed on the device. Because EXO-UL7 is a redundant manipulator and supports the entire configuration of the arm, a single force sensor on the device end effector is not enough to fully define the movement. Additional force sensors on the device that are located at each attachment point of the machine interface allows for the full configuration of the device to be specified. This turns the under defined problem into an over defined problem (too many force signals for the number of Dof). Two strategies are developed to project the signals onto a seven degree subspace. The first adds the force vectors in task space then uses the Inverse kinematics to develop joint trajectories. The second uses the Jacobian transpose to map the forces first to instantaneous joint torques. Then the torques are added in joint space and finally joint trajectories are developed from joint torques. Six subjects performed a peg in hole experiment and it was found that task based admittance control had about 11% lower interaction energy required to do the task. It was also shown that with both admittance controllers the subjects did the tasks slower than with no control at all. Kinematic and dynamic constraints reduced the bandwidth of the system. To improve the bandwidth, predictive algorithms are employed.;Swivel angle prediction is the first predictive algorithm implements to improve the performance of the device. With this control strategy the configuration of the redundant space is related to the end effector position. by observing human behavior it was noted that one of the fundamental task preformed by the human arm was to bring food to the mouth. By maximizing the manipulability of the device in the direction of the mouth, a simple stable closed form prediction of the configuration of the device was achieved. Comparing movement from motion capture to predicted motion showed a good correlation and testing of the algorithm on the exoskeleton device was conducted. 4 subjects conducted a peg in hole task. An 11.22% reduction of interaction energy was achieved when compared to Admittance control alone. This algorithm can be used for motion following as in the current set up, or to predict what the arm configuration should be when working with disabled populations. Although this method predicted motion very well, it only provided prediction of the one Dof redundant space of the arm.;To further extend the prediction capability and motion following of the device, neural control was implemented in which electro-myography (EMG) is used to read the nerve impulses to the muscle. Although this signal only relates muscle force to isometric muscle contraction, using other system parameters read from Exo-UL7 such as the joint, position and velocity, a Hill based muscle model predicts the muscle force and ultimately the muscle torque. Due to an inherent delay between when the nerve impulse can first be detected and when the muscle contracts (somewhere on the order of 50-100 ms) the motion can be predicted before the arm begins to move. The model has many variables that need to be predicted for each individual so before each subject test an parameter fitting is conducted. Four subjects preformed a peg-in-hole test. It was shown that the interaction power increased compared to admittance control, but the completion time decreased. With further examination it was noted that the interaction force and energy when using the neural control was the same as with the admittance control. This implies that with the same force we achieved a higher velocity, which means that the system had a higher overall gain. The performance gains were not uniform through out all the subjects. The parameter fit conducted for each subject did not guarantee convergence to even a local minimum and there are still opportunities to improve the system performance. (Abstract shortened by UMI.).