Sensor motion planning with active uncertainty reduction: gripping and docking tasks

This paper deals with planning of motion for a robot approaching a box shaped object using a range sensor. The plan is given as the minimizing solution of a criterion approximating the expectation of the quadratic error for the final position. The expectation is calculated using Gaussian sum approximations for the distributions of the stochastic variables involved. An approximation of the information received by a range sensor measuring a box is also presented. The resulting criterion is minimized over the sensor positions by a gradient descent algorithm yielding sub optimal solutions. Even though the solutions are only sub optimal they display both probing and cautious behavior reflecting the uncertainties involved. As new observations become available the future sensor positions are replanned with the horizon reduced by one step, adjusting the solution with the updated estimated position and covariance