UC San Diego
Exploratory studies of human sensorimotor learning with system identification and stochastic optimal control
- Author(s): Simpkins, Charles Alexander
- et al.
Biological sensorimotor control systems possess the ability to achieve control objectives under circumstances which would challenge even the most masterful control engineer - high dimensionality, noise, redundancy, uncertainty, continuously changing tasks, and delays. Even the control objective itself may be uncertain and require some exploration to determine. Therefore by studying these systems, science and engineering can benefit from the knowledge that will inevitably be gained. Stochastic optimal control has been successfully used to model human sensorimotor behavior in many contexts, and, combined with Bayesian inference, the framework can be extended to address problems where sensorimotor learning must take place. This dissertation contributes a suitable model for sensorimotor learning applicable to a broad class of problems (including, high dimensional, nonlinear, and stochastic contexts). The second contribution is solution methods based on function approximation and techniques for practical application of these methods for producing globally (approximately) optimal controllers. These controllers are compared (favorably) to human subjects performing exploration/exploitation tasks involving redundancy and uncertainty. The active exploration framework developed here can be applied to artificial systems for creating systems capable of anticipating what a task is and how to achieve the goals. The third contribution is robotic designs for testing, studying and simulating sensorimotor learning and control. Preliminary results demonstrate that the designs succeed in possessing the required capabilities for mimicking biological systems. This design methodology is highly effective regarding modularity, backdrivability, compliance, and responsiveness. Previous studies have shown that the sensorimotor system simplifies control by combining individual muscle activations into synergies, allowing fewer high level ̀control knobs' mixed into complex behaviors. This notion of 'synergy' is redefined to include visual and motor synergies. A measure of dimensionality of synergies and variability of the sensorimotor system is created by presenting subjects with a stereoscopic image of a complex hand posture for them to match while measuring their hand postures, trajectories from a nominal posture, and repetitions. Methods are also presented to examine the sources of motor variability. Results differ from other recent work. Task difficulty is found to affect the number of synergies available