| Description |
The goals of this dissertation are twofold: (1) to overcome the limitations of existing manufacturing processes for creating three-dimensional (3D) ionic polymer-metal composite (IPMC) soft actuators and sensors by developing a new free-form additive manufacturing (3D printing) process, and (2) to address the challenges of motion control of fabricated IPMC-based actuators through degradation modeling and machine learning control, specifically leveraging Bayesian-based methods for performance projections and control policy search. IPMCs and other ionic polymer conductor composites are kinds of electroactive polymers that can be used as soft actuators and sensors. In particular, IPMCs consist of an ion-exchange membrane (a polyelectrolyte) and outer surface electrodes, where application of a voltage across the electrodes of a hydrated IPMC causes the composite to deform, and when deformed, the IPMC produces a millivolt-scale electric signal. Thus, IPMCs can function as actuators and sensors. As such, IPMCs are functional down to sub-micron scales, can exhibit large displacements, are controlled through low voltages (<3 volts), operate in hydrated environments, and are exible. However, two main challenges to utilizing IPMCs in practical applications are that there are not well-developed freeform fabrication methods to create monolithic, custom-shaped IPMCs and IPMCs exhibit idiosyncratic, time-varying, and complex dynamic behavior (including back-relaxation), which makes motion control dicult. More speci cally, two IPMCs fabricated through the same process will not exhibit the same behavior; behavior of a single IPMC will vary over time; even the idealized electromechanical behavior of an IPMC is determined by nonlinear elasticity and impedance, and must account for phenomena such as back-relaxation and high-order resonances to be accurate. In addition to this, their performance degrades when operated above the decomposition voltage of the hydrating solvent. The rst contribution of this dissertation is a new fused deposition modeling (FDM) 3D printing process developed to create custom-shaped, non-planar IPMCs where the key advance is adaptation of precursor of the ionomeric material, Na on, to a melt-processing-based 3D printing method. This process requires in-house fabrication of precursor lament and custom modi cations to a conventional FDM 3D printer that enables e ective printing at elevated temperatures (between 100 C and 200 C for the build stage and between 280 C and 300 C for the print head). The new IPMC 3D-printing method is used to create example electroactive polymer devices, such as a linear actuator, a rotary actuator, a multi-degree-of-freedom actuator, a gripper, and a soft crawling robot. The second contribution is characterization of IPMC performance degradation in terms of accepted electromechanical models and a method for projecting performance degradation from past experiments and real-time data to account for IPMC performance degradation with extended use. Characterization of 3D-printed IPMC actuators shows that, when operated continuously over an 8-hour period, all IPMCs exhibit decreased sti ness, and that all IPMCs operated at 3-volt amplitudes (1.75 volts above the electrolysis voltage of water) also exhibit an increase in resistance and a decrease in the electromechanical transduction coecient by at least an order of magnitude. To account for performance degradation, a Gaussian process is used to model IPMC performance degradation based on the time integral of the input to the IPMC actuator, the planned future inputs to the IPMC actuator, and the IPMC actuator's initial performance conditions. The Gaussian process is trained on the data for all tested IPMC actuators. This approach results in an overall mean prediction error of 10.78% of the 2.5-mm range when conducting leave-one-out cross-validation for each tested IPMC actuator. Bayesian updating is then used to adjust model projections based on the real-time performance of the particular IPMC actuator. This reduces the mean error to 6.64% of the 2.5-mm range. Planning for performance degradation using predictions from the Gaussian process that is adjusted based on real-time performance data is found to reduce the cumulative squared under-performance by as much as 50% over control methods that do not account for performance degradation. Finally, the third contribution of this dissertation is an investigation of Bayesian optimization as a fast-converging policy search method to address the challenges to motion control of IPMC-based devices. This approach makes use of incremental measurements of a relevant performance metric to converge to an e ective control policy for devices used in a repetitive task. For example, the control policy that dictates the gait of the crawling robot is tuned based on the distance the robot travels. Simulations and experimental results demonstrate that using a known reachable performance level explicitly as the target in the acquisition function, during Bayesian optimization, signi cantly increases the rate of convergence. However, it is also concluded that performance degradation of IPMC-based devices is signi cant enough, when operated above the electrolysis voltage of the hydrating solvent, to make the target performance level unreachable if the performance degradation is not accounted for. |
