![](bg6.png)
Table 2: Performance indexes of gravity compensation
models.
ANN compensation
(on-field training)
ANN compensation
(simulation training)
Analytical model
compensation
7 CONCLUSIONS
The aim of this work was to investigate the
performance of ANN’s for the gravity compensation
of a spherical parallel manipulator. The availability
of an analytical model of robot’s dynamics allowed
to design the net through computer simulation and to
port the software to the real-time control hardware
after a first (preliminary) tuning.
After the on-field net training phase, more tests
have been performed to evaluate the compensation
effectiveness: the manipulator has been commanded
to reach several poses in the task space and the
output of the ANN has been compared with the
actuation forces actually needed; the same
comparison has been made with the actuation forces
provided by the analytical model of robot’s inverse
dynamics.
The results of the tests showed that the ANN
controller behaves better than the algorithmic one,
even if the improvements are not very relevant. It is
reckoned that better performances could be obtained
by a more extensive training on the field and in case
the mechanical structure were less influenced by
static friction, as is the present case: the influence of
this factor could be assessed by further
experimentation, by measuring the gravitational
actions with the manipulator in (slow) motion.
From the computational point of view, on the
other hand, the ANN based compensation
overperforms the analytical model, whose
computation is rather time consuming indeed.
REFERENCES
Callegari, M., Carbonari, L., Palmieri, G. and Palpacelli,
M.-C. (2012). Parallel Wrists for Enhancing Grasping
Performances. In: Grasping in Robotics, G. Carbone
ed., Springer, 2012. (In press)
Callegari, M., Marzetti, P., Olivieri, B. (2004). Kinematics
of a Parallel Mechanism for the Generation of
Spherical Motions, in J. Lenarcic and C. Galletti (eds),
On Advances in Robot Kinematics, Kluwer, pp.449-
458.
Carbonari, L. (2012). Extended analysis of the 3-CPU
reconfigurable class of parallel robotic manipulators.
Ph.D. Thesis, Polytechnic University of Marche,
Ancona, Italy
Cybenco, G. (1989). Approximation by Superpositions of
a Sigmoidal Function. Mathematics Control, Signals
and Systems, Vol. 4, no. 2, pp. 303-314.
Hagan, M.T. & Menhai, M. (1994). Training feed-forward
networks with the Marquardt algorithm, IEEE
Transactions on Neural Networks, Vol.5, No. 6,
pp.989-993
King, S.-Y., and Hwang, J. -N. (1989). Neural network
architectures for robotic applications. IEEE
Transactions on Robotics and Automation, Vol. 5, no.
5, pp. 641-657.
Kwan, H. K. (D. of E. E.). (1992). Simple sigmoid-like
activation function for digital hardware
implementation. Electronics Letters, Vol. 28, no. 15,
pp. 1379-1380.
Narendra, K. S. and Parthasarathy, K. (1990).
Identification and control of dynamical systems using
neural networks, IEEE Transaction on Neural
Networks, Vol. 1, no. 1, pp. 4-27.
Orlowska-Kowalska, K and Kaminski, M, (2011). FPGA
Implementation of the Multilayer Neural Network for
speed estimation of the Two-Mass Drive System.
IEEE Transaction on Industrial Informatics, Vol. 7,
no.3, pp. 436–445.
Serrano, F. E. (2011). Neural Networks Models for
Control and Identification of a Two Links Robotics
Manipulator. Proceedings CONCAPAN XXXI.
Zhang, D., & Lei, J. (2011). Kinematic analysis of a novel
3-DOF actuation redundant parallel manipulator using
artificial intelligence approach. Robotics and
Computer-Integrated Manufacturing, Vol. 27, no. 1,
pp. 157-163.
Design and Experimentation of a Neural Network Controller for a Spherical Parallel Robot
255