UCLA

This thesis develops adaptive and iterative learning control methods for robot trajectory tracking applications. Specifically, iterative learning control is applied when a desired reference trajectory is known beforehand, and adaptive control is designed to cope with unknown patient motion (disturbance) in MRI-guided robot-assisted intervention.

For robot manipulator tracking, a nested-loop iterative learning control is proposed. This method requires only nominal kinematic parameters from factory setting, gives fast convergence, and can be added on top of existing servo loop. The ILC learning architecture includes an inner loop that accounts for motor dynamics, and an outer loop that addresses the static bias from the payload or imprecise kinematics. A data-based learning filter design is extended to cope with motion constraint and multivariate systems. It is experimentally verified on a 6-DOF serial robot that the proposed method mitigates the maximum dynamic tracking error by an order of magnitude, and is applicable to different payloads due to small system variation from torque shielding of gear reduction.

For tracking of general nonlinear dynamic systems, an efficient data-driven ILC algorithm is proposed. As opposed to the model-based methods, for which nonlinear identification and learning law design can be cumbersome, this method uses adaptive filter to implicitly (and automatically) construct linearized system inverse for effective learning. An existing adjoint-based ILC for LTI system is also extended to cope with nonlinear dynamics, and for comparative study. The SISO algorithms are simulated and experimentally validated on a fully-actuated 2-DOF laboratory pendulum system. Algorithms are also developed to circumvent the difficulty when adapting a right inverse for MIMO systems.

The automated MRI-guided intervention is motivated by the current procedural inefficiency from constraints posed by MR environment. As lots of researchers focus on either MR-safe/conditional robot to augment the reach of the physician, or MR image tracking for motion estimation of tissue/instrument, this work aims at addressing a more flexible setting: use real-time MRI for instrument control when a target is in motion. It is enabled via the integration of robot hardware, MRI sensing, and control techniques. On the control aspect, we characterize the MR imaging process and the robot dynamics, then propose adaptive control schemes to overcome the long delay and high noise variance from MRI measurement. The study is conducted on a hydrostatically actuated platform, which consists of a target motion module that emulates respiratory motion, and an instrument manipulation module regulating the instrument-target distance.