Neuromuscular stochastic optimal control of a tendon driven index finger model

Our long-term goal is to find control principles to control robotic hands with dexterity and robustness comparable to that of the human hand. Here we explore a control strategy capable of accommodating the nonlinearities, high dimensionality and endogenous noise intrinsic to complex, tendon-driven biomechanical structures. We present the first stochastic optimal feedback controller (i.e., an iterative Linear Quadratic Gaussian controller) applied to a tendon-driven simulated robotic index finger model. In our model we take into account both the tendon network driving of the index finger, and we consider first-order muscle dynamics. Our feedback controller shows robustness against noise and perturbation of the dynamics. Moreover, it can also successfully overcome the nonlinearities intrinsic to the mechanics of the finger for large postural changes, and the need for non-negative control signals. Our simulations provide, for the first time, the complete time history of tendon tensions, lengths and velocities for the tasks of tapping with nonzero terminal velocities required for dynamic manipulation. We find that the optimal control of realistic tendon-driven systems fundamentally stretches current methods to their limits. To find a successful control strategy, the algorithm must overcome several critical challenges inherent to the control of tendon-driven fingers systems in which all uni-directional control commands can actuate all joints (either directly or through dynamic coupling). Therefore, all elements of the solution are interwoven including the tuning of the cost function, the dynamics of the plant, and the initial guesses for state and control trajectories.