Neural Network Development and Training for the Simulation of Dynamic Robot Movement Behavior
In this chapter the design and evaluation of artificial neural networks for learning static and dynamic positioning behavior of an industrial robot are presented. For the collection of training data, an approach based on the Levenberg–Marquardt algorithm
- PDF / 2,075,054 Bytes
- 13 Pages / 439.37 x 666.142 pts Page_size
- 94 Downloads / 194 Views
Abstract In this chapter the design and evaluation of artificial neural networks for learning static and dynamic positioning behavior of an industrial robot are presented. For the collection of training data, an approach based on the Levenberg–Marquardt algorithm was used to calibrate the robot and the coordinate measuring machine to a common reference system. A sequential approach for the network design development is presented. The network was verified by measuring different robot path segments with varying motion parameters, e.g. speed, payload and path geometry. Different layouts and configurations of feed-forward networks with backpropagation learning algorithms were examined resulting in a multi-layer network based on the calculation of the forward transformation.
1 Introduction Industrial robots come with their own set of coordinate systems to express Cartesian positions in their domain. A typical set includes definitions for world, tool and work object frames. These frames describe the position and orientation (pose) of the robot Based on Learning Robot Behavior with Artificial Neural Networks and a Coordinate Measuring Machine, by Benjamin Johnen, Carsten Scheele and Bernd Kuhlenkötter which appeared in the Proceedings of the 5th International Conference on Automation, Robotics and Applications (ICARA 2011). Ó 2011 IEEE. B. Johnen (&) C. Scheele B. Kuhlenkötter Industrial Robotics and Production Automation (IRPA), TU Dortmund University, Dortmund, Germany e-mail: [email protected] C. Scheele e-mail: [email protected] B. Kuhlenkötter e-mail: [email protected]
G. Sen Gupta et al. (eds.), Recent Advances in Robotics and Automation, Studies in Computational Intelligence 480, DOI: 10.1007/978-3-642-37387-9_1, Ó Springer-Verlag Berlin Heidelberg 2013
1
2
B. Johnen et al.
in the overall setup, the pose of the tool and the pose of the work object. The robot establishes its own TCP pose by solving the forward kinematics problem: This calculation converts the representation of the TCP pose in joint space (which is e.g. described by a vector of the current values of all joints) into the Cartesian space (cf. [1]). This approach is prone to inaccuracies from various factors. In the static case, these factors are e.g. transmission tolerances and fabrication tolerances such as arm length. When the robot is moving (dynamic case), positioning errors may arise from joint friction or robot payload (cf. [2]). For an impartial acquisition of these inaccuracies, an external measurement system is required. The precise measurement of the robot deviation opens up the possibility to apply learning methods for creating a robot simulation that will be able to predict these deviations. In this chapter, we present an approach where we design and train artificial neural networks to build static and dynamic models of the robot behavior. The robot used in the experiments is an ABB IRB 2400/16 (handling weight 16 kg, see Sect. 2); the coordinate measuring machine is a Nikon Metrology K61
Data Loading...
