Abstract
Humans frequently and successfully reach to objects without precise information about their location. This success suggests humans have developed motor strategies that compensate for location uncertainty. To investigate this issue, we conducted experiments that involved grasping and lifting cylinders whose position was not precisely known. Position uncertainty was introduced by randomly moving the cylinder with a robotic arm over sequence of 5 positions sampled from a strongly oriented 2-D Gaussian distribution. Preceding the reach, vision of the object was removed using liquid crystal shutter glasses, and the robot moved the object one additional time. Participants reached and grasped the object while out of view. Finger trajectories were recorded for a set of covariance orientations. Preliminary results show that human grasping compensates for position uncertainty of objects. All subjects follow a similar strategy so that to maximize the probability of contact with the object: The approach angle increases almost linearly with the covariance angle, so that the fingers approach the cylinder in the direction of maximal cylinder uncertainty. Moreover, the finger and thumb slow down and become parallel as they approach the object. To interpret these results, we have analyzed the data using methods from reinforcement learning and optimal control. In particular we have used the experimental data to learn human policies and cost-to-go functions for grasping objects with position uncertainty. We will present a detailed comparison of human strategies with optimal policies to test the optimality of human compensation strategies. In addition, we investigated the effects of the 5equential cylinder's positions to the finger and thumb trajectory using regression analysis to address whether observers use information from the 5 movements to estimate the cylinder's position on each trial.
NEI R01 EY015261, ONR N00014-05-1-0124.