On Robot Feedback from Range Sensors : Reliable Control by Active Reduction of Uncertainty and Ambiguities

Abstract: This thesis is on modelling and experimental tests when non contact sensing isused for feedback control in robotics. The motion of the robot is to be controlled relative to objects in the surrounding workspace during operations like gripping/docking, surface following, shape measuring etc.On a system level three examples are studied in detail:• Tolerance driven docking of a mobile robot to a rectangular box. Experiments with feedback from a range camera are presented. The control law is a LQG-design which adapts the trajectory and the speed to fulfil given posture tolerances at the docking position.• Eye-in-hand control of a flexible arm using a range sensor for direct measurement of the tip-to-object distance. In simulations, vibrations are reduced by feedback, improving the settling time with at least a factor five.• Navigation of a robot using telecommands, gives an example of a framework where feedback sub-tasks are integrated as telecommands.The long term goal is a framework where sub-operations can be assembled using model based feedback designed to fulfil the task tolerances with a pre-specified probability (close to unity). In essence, the tolerances and the required probabilityof success should be a model parameter of the feedback design.Estimation is an important part of the presented feedback loops and implications are studied for four cases:• Specular reflections often results in false peaks for a sheet-of-light range camera. Models for detecting and reducing such ambiguities are studied.• An algorithm for position estimation of an object with initially large orientation uncertainty is presented. Gaussian mixtures are used to represent the uncertainty in orientation.• Motion planning to maximise information from a scanning ladar range sensor during approach towards a rectangular object. A criterion based on the expected final position error is minimised over a given number of future sensing positions during the approach.• Modelling of covariances for planes and intersections estimated from range measurements and straight edges in vision. The covariances are parametrised in a form where effects of robot motion and the center of gravity for the measurements is apparent.