Vision-based robot path planning

Abstract-- We present a vision-based approach to robot path planning. The programmer demonstrates a desired motion by moving an object to be manipulated with his own hand. His performance is measured with a stereo vision system. The demonstrated motion is

In: Advances in Robot Kinematics and Computational Geometry, J. Lenarcic and B. Ravani, eds., pp. 505-512, Kluwer, Dordrecht, 1994Vision-Based Robot Path Planning

Ales Ude

Rudiger Dillmann

Institute for Real-Time Computer Systems and Robotics, Department of Computer Science, University of Karlsruhe, Kaiserstr. 12, 76128 Karlsruhe, Germany

Abstract{ We present a vision-based approach to robot path planning.

The programmer demonstrates a desired motion by moving an object to be manipulated with his own hand. His performance is measured with a stereo vision system. The demonstrated motion is reconstructed with the help of a non-parametric regression technique which is resistant to a considerable amount of measurement noise. The generated path is given as a linear combination of natural vector splines. The proposed methodology requires neither a geometrical model of a robot workcell, which is used by classical motion planners, nor a robot manipulator, which is needed for teaching by guiding, for the speci cation of the robot path.

The most widespread approach to robot path speci cation in the industry is teaching by guiding which involves guiding the robot through a sequence of poses which are recorded by the robot's internal sensors. This simple method, however, su ers from a number of di culties. It forces the user to use a teach-pendant, which is tiring and cumbersome and often leads to mistakes in the movements, especially if ne positioning is needed. Furthermore, the use of real manipulators can be dangerous for the programmer. Many recent approaches to robot path speci cation fall in the category of geometrical motion planning. The basic motion planning problem is de ned as follows: given an initial pose and a goal pose, generate a path specifying a continuous sequence of poses avoiding contact with the obstacles in the robot workspace, starting at the initial pose and terminating at the goal pose. There exist many extensions to the basic motion planning problem 1]. Classical motion planning algorithms su er from computational complexity and dependency on the model of the robot workspace. Their conversion into commercial systems has turned out to be an extremely di cult task. There is an adjustment problem due to inconsistencies between the model and the real environment and there are many problems in which the robot's environment is only partially known. The generation of geometrical models is usually carried out with the help of CAD systems and is a tedious task. When the model of the robot workcell is not available, classical motion planning algorithms cannot be used. The robot is forced to use local obstacle avoidance strategies based on on-line sensing to move from its initial to its nal con guration. However, on-line sensing is computationally expensive and often cannot be performed in real-time. Furthermore, the lack of global knowledge results in non-optimal motions. Obstacle avoidance is not always the only criterion which the

robot path must ful ll. Especially in adaptive manufacturing processes such as spray painting, the robot end-e ector is required to follow a well de ned path in the robot workspace. In this case, the primary goal

I Introduction

Abstract-- We present a vision-based approach to robot path planning. The programmer demonstrates a desired motion by moving an object to be manipulated with his own hand. His performance is measured with a stereo vision system. The demonstrated motion is

of the planner is to specify such path. It is di cult to tackle this kind of problems with the help of classical motion planners because every task must ful ll its own criterion. Di erent motion planners must be used for di erent tasks as a consequence. Another approach is to exploit the human programmer's implicit knowledge about the path to be programmed. The programmer usually has some idea about the course of the desired robot motion but lacks of any means to communicate this path to the robot. A CAD oriented trajectory design-editor was proposed to support the explicit programming of robot paths 2], but such an editor again depends on the geometrical model of the workcell. In this paper we employ a teaching by showing paradigm 3] as an e cient approach to explicit programming of robot paths. Methods for the speci cation of robot poses and paths which utilize teaching by showing were proposed in 4, 5, 6, 7]. In 5] vision was employed to teach the poses on the desired path at some important passing points whereas in 4, 6] vision was used to specify overall paths. In these systems, the path speci cation was accomplished by moving a specially designed teaching tool with attached LEDs along the desired path. Various heuristics were developed in order to reduce the noise in it. However, no attempt was made to assure the optimality of the reconstructed path according to a suitable optimization criterion and to explicitly consider the noise of the sensors. By designing a tool which can be easily recognized by the image processing system, the object tracking can be made faster and more robust. But no teaching tool can be appropriate to show all possible paths and it is sometimes easier to show the desired path by moving the actual object which should be manipulated. The inability to track more general objects is a serious disadvantage of both methods. In 7, 8] we described a theoretically well supported approach to programming of robot trajectories based on teaching by showing. The user demonstrates the motion by moving the actual object to be manipulated along the desired path. The proposed methodology employs a non-parametric regression technique …… 此处隐藏：19536字，全部文档内容请下载后查看。喜欢就下载吧 ……