Multisensor-based robotic manipulation in uncalibrated environments.

摘要

This dissertation is aimed at planning and control issues of robotic manipulation in uncalibrated environments by using multisensor fusion schemes. Two different but typical tasks in manufacturing engineering are considered in a common framework. One is to control a robot to track, grasp and pick up a moving part in a reconfigurable workcell. It is called the tracking and grasping task. The other, referred to as the trajectory-following task, is to drive a tool grasped by the robot to follow a visible trajectory on an unknown surface. The latter is actually a new formulation of a typical task in industry. Many tasks in manufacturing engineering such as moving objects from one place to another, painting, welding and cutting materials along a certain path, can be classified into the categories. Difficulties arise when we assume that the robot works in uncalibrated environments with an uncalibrated camera. To work successfully in the uncalibrated environments, multiple sensors are employed. They are encoders mounted on each joint of the robot and the motor of the conveyor, the camera fixed above the workcell and a force/torque sensor mounted on the wrist of the robot. The whole system under our consideration consists of several subsystems: the robot manipulator, the vision system and the conveyor (or unknown surface for the trajectory-following task). Novel multisensor fusion schemes are developed for planning and control of the robot without knowledge of the relative pose between the subsystems. For the tracking and grasping task, we utilize a virtual rotation algorithm to transform original image data into "top-view" information. A multi-image approach is provided to determine points on a non-planar part by using the single camera which is uncalibrated with respect to the workcell. By solving an optimization problem, we can determine the relationship between the fixed disc frame and the base frame of the robot based on sensor fusion schemes. The trajectory of the moving part is obtained in real-time by means of sensor fusion. As a result, the desired trajectory for the robot can be generated for control purpose. As far as the trajectory-following task is concerned, only good estimate of the trajectory is not enough to assure the completion of the task. The hybrid position/force control strategy is adopted and a new hybrid control law is derived based on sensory information by decoupling the control variables into two parts. Motion planning is done in accordance with the visual information and the measurements from encoders of the robot and the force/torque sensor. The combination of the information from the force/torque sensor and the vision system guarantees the completion of the task. Simulations and experiments are carried out to verify the feasibility of our proposed methods for the two tasks. The advantages of our proposed approaches include: (i) the requirements of the computation speed for the vision system is greatly weaken; (ii) the whole system has flexibility and intelligence in the sense that it can work in re-configurable and uncalibrated environments without explicit intervention or reprogramming.