Improvements and conclusions

As outlined in the previous section, the experiments com­pleted so far make use of a very crude and limited rig. However, the success attained demonstrates the feasibil­ity and the potential of the approach. The nature of the x-y table used means that the kinematics of the system are trivial, and that dynamics do not come into consid­eration. The use of stepper motors alleviates the need for closed loop joint control. The addition of a second camera would circumvent any need for calibration, since the ambiguity present in a single image could be re­moved. It may also prove beneficial to incorporate other sensors to provide additional information, such as force feedback [17]. Integration of data from multiple sensors can improve robustness; any redundancy in the informa­tion provided can be used to highlight noise and spurious sensor readings [18].

The learning system presented here is still under devel­opment and could be improved in many ways. Different algorithms for generalisation to new situations and man­agement of the potentially large number of experiences are currently under consideration. More extensive and demanding tests are required to prove that the approach is suitable for use in an industrial setting.

The use of visual feedback as a means to compensate for the poor mechanics of a cheap robot has been demon­strated. Due to the nature of this approach, the system is flexible with respect to the position of the workpiece. Initial experiments indicate that a learning controller is capable of adapting to the behaviour of a specific robot. Effects due to the low cost nature of the robot, such as those due to backlash, can be overcome.

Work is underway to extend the ideas presented to PCB component placement. The majority of industrial PCB assembly machines are built with speed of operation a critical consideration, and therefore do not provide a real­istic comparison in terms of cost. However, one commer­cially available component placement machine of compar­able speed [19] sells for over 120,000. This demonstrates the potential cost savings of the approach outlined in this paper. The basic principles applied should readily extend to a variety of other robotic applications, thereby making them considerably more cost effective.

REFERENCES

1. GROOVER, M.P., WEISS, М., NAGEL, R.N. and ODREY, N.G., Industrial robotics: Technology, pro­gramming and applications, McGraw-Hill, 1986.

2. PLATT, N., ‘Robots in industry: An overview of robot applications’, Cambridge University Engineering Depart­ment Seminar, ABB Robotics, 1994.

3. CORKE, P.I., ‘Visual control of robot manipulators - a review’, in HASHIMOTO, K., editor, ‘Visual Servoing’, 1-32, World Scientific, 1993.

4. WI.JESOMA, S.W., WOLFE, D.F.H. and RICHARDS, R.J., ‘Eye-to-hand coordination for vision-guided robot control applications’, Int. -J. Robotics Research. 65-78, volume 12, number 1, 1993.

5. К ASS, М., WITKIN, A. and TERZOPOULOS, !).. ‘Snakes: Active contour models’, Proceedings of the 1st Intl. Conf. on Computer Vision, 259-268, 1987.

6. HODGES, S.E. and RICHARDS, R.J., ‘Fast multi­resolution image processing for PCB manufacture’, IEE Colloquium on ‘Multi-resolution modelling and analysis in image processing and computer vision’, Digest No. 1995/077, 8/1-8, 1995.

7. HODGES, S.E. and RICHARDS, R.J., ‘Uncalibrated stereo vision for PCB manufacture’, IEE Colloquium on ‘Applications of machine vision’, Digest No. 1995/113, 4/1-6, 1995.

8. MUNDY, and ZISSERMAN, A., (editors), ‘Geo­metric invariance in computer vision’, MIT Press, 1992.

9. MARTINEZ, T.M., RITTER, H.J. and SCHUL- TEN, K.J., ‘Three-Dimensional Neural Net for Learning Visuomotor Coordination of a Robot Arm’, IEEE Transactions on Neural Networks. 131-136, volume 1, number 1, 1990.

10. NGUYEN, I... PATEL, R.V. and KHORASANI, K., ‘Neural Network Architectures for the Forward Kinemat­ics Problem in Robotics’, Proc. of IEEE Intl. Conf. on Neural Networks, 393-399, Volume III, 1989.

11. MOORE, A.W., ‘Efficient Memory-Based Learning for Robot Control’, Technical Report No. 209, Computer Laboratory, University of Cambridge, 1990.

12. OMOHUNDRO, S.М., ‘Efficient Algorithms with Neural Network Behavior’, Report No. UIUCDCS-R- 87-1331, Department of Computer Science, University of Illinois at Urbana-Champaign, 1987.

13. HENNESSY, J.L., and PATTERSON, D.A., ‘Com­puter architecture: A quantitative approach’, Morgan Kaufmann, 1990.

14. VLSI VISION LTD., ‘VVL1070 provisional data­sheet’, VLSI Vision Ltd, Edinburgh, 1994.

15. ANDERSON, L., (editor), ‘Achieving assembly flex­ibility for printed circuit boards’, Machine Vision World. 20-21, volume 7, number 6, 1989.

16. McVEY, E.S. and Van TOL, A., ‘An Experi­mental Printed Circuit Board Drilling System Automated By Pattern Recognition’, Pattern Recognition, 271-276, volume 11, 1979.

17. SELKE, K.K.W., ‘Generic assembly applied to indus­trial products’, Industrial Robot. 7-10, volume 20, num­ber 3, 1993.

18. HODGES, S. and LOUIE, G., ‘Towards the interact­ive office’, Human Factors in Computing Systems, CHI Conference Companion, 305-306, ACM Press, 1994.

19. VERSATRONICS, ‘The R-V Placer datasheet’, Ver- satronics Ltd., Witney, Oxon, 1994.