Robotics Research in NAIST

Learning Oculomotor Control

Theoretical Life-Science, ATR, and JST.

This is a interdisciplinary research between computatinal neuroscience and robot vision, focusing on biomimetic oculomotor system. The binocular vision system mimics primates, having foveal and peripheral vision which are realized by the combination of a zoom-lensed camera and a wide-view camera, respectively.

This oculomotor control system shares the similar computational problems with humans: (1) its dynamics is unknown, (2) there is 0.1 s delay in the visual feedback pathway, (3) real-time perception is required in the real-word contaminated by the non-Gaussian noise, and (4) the foveal and peripheral visions should be synthesized in some effective manner. We have already proposed algorithms for (1) and (2) by means of on-line dynamics learning of the self and the external world. We also have proposed a novel particle finterling algorithm for (3). (4) is a problem we will work on.

Humanoid Robot DB（in ATR）

in an experiment foveated vision

Reference

Bando, T., Shibata, T., Doya, K., and Ishii, S. Switching Particle Filters for Efficient Real-time Visual Tracking. Int Conf Pat Recog (ICPR2004), Cambridge, U.K., 2004 (in press).
Shibata, T. and Vijayakumar, S. and Conradt, J. and Schaal, S. Biomimetic Oculomotor Control. Adaptive Behavior, 9 (3/4), 2001. pp.189-208.
Shibata, T. and Schaal, S. Biomimetic Gaze Stabilization based on Feedback-Error-Learning with Nonparametric Regression Networks. Neural Networks, 14(2), 2001, pp.201-216.