The representation, learning, and control of dexterous motor skills in humans and humanoid robots

摘要

Humans are capable of executing a wide variety of complex and dexterous motor behavior with grace and elegance. We would like to understand the computational principals underlying their sensorimotor control, both for the advancement of neuroscience, as well as to achieve similar performance on artificial machines. I study the problem via two approaches: motor control experiments on human subjects, and the implementation of control models on anthropomorphic hardware. In one experiment, we use a novel exoskeleton platform that permits us to alter arm dynamics by applying torques to specific joints during full 3-D motion, and study how arm redundancy is utilized. Results of this experiments suggest that: 1) humans are able to learn, and later predict, novel dynamic environments, 2) humans plan motion in an extrinsic hand, or end-effector oriented space, and 3) joint redundancy can be utilized in order to assure task achievement. In a second experiment, we attempt to create a dynamic environment that effects motor cost, but not accuracy: if a force perturbation first pushes the hand off-course in one direction and then subsequently in the opposite direction, the reaching task may still be achieved with minimal correction, but a strongly curved trajectory. Under such conditions, subjects learn to return their trajectories towards the baseline, or null field trajectory, effectively fighting this disturbance in order to maintain a straight trajectory. These results are inconsistent with theories suggesting that perceived kinematic error plays no role in motor adaptation. I will demonstrate how improved predictions can be obtained with a stochastic optimal controller using a directionally biased cost function. In second part of my dissertation, I discuss how to obtain robust, compliant, human-like motion of humanoid robots, via task-space, model-based, feed-forward control.