Kurt3D facing a box

Kurt3D facing a cylinder

Kurt3D facing a gap

CONCLUSIONS:

In this study, traversability affordances of the environment for a mobile robot is learned through physical interactions in a physics based simulation environment. Since the traversability depends on the location of the objects and their geometrical properties, range images are used to perceive the physical affordances of the immediate environment. A simple perceptual representation is proposed, where intermediate high-level processes like object detection or world modeling are not utilized, thus favoring Gibsonian direct perception view. Based on the low-level features that are perceived and the results of the interactions with the world, the robot is able to learn i) relevant features for different actions, and ii) the affordances provided. The prediction accuracy in perceiving the traversability affordances of the environment, which include several boxes, cylinders, and spheres is found to be around 95%. Furthermore, it is presented that the robot uses only 1.1% of the extracted features while perceiving the affordances. This in turn save the time 76.6% in scanning and 99% in feature processing, and J.J. Gibson's perceptual economy is obtained through learning to use relevant features.

After learning the affordances of the environment, the robot is tested in various setups. It is placed in a virtual cluttered room, and controlled with a simple motivation system. In this environment, the robot was able to traverse the environment, by successfully selecting its actions based on the perceived affordances. In the next experiment set, the generalization performance of the learned affordance based perception system is analyzed. It is shown that the robot was able to perceive the traversability affordances of the novel objects that it has never seen before. In the last set of experiments, the affordance-based action selection scheme that is learned in simulator is successfully transfered to real robot without any further modification. Although there is no concept of object or width in any representation level, and the robot has no awareness of its own body dimensions, it is able to perceive the traversability affordances of the apertures between the objects. In other words, the affordances of the apertures, which depend on the relation between the width of the apertures and the shoulder width of the robot, are directly perceived without recognizing them.

The work presented in this paper is novel from prior studies on multiple fronts. First, in our work range images, which are more informative about the physical affordances of the environment, are used for sensing. Second, we proposed a perceptual representation which represents the shape and orientation information in a proper way for learning. Third, we performed a more complete and comprehensive testing of the learned affordances, and that showed that the proposed system can successfully predict the affordances of completely novel object types.

Bibliography

1

W. Warren, ``Perceiving affordances: Visual guidance of stair climbing,'' Journal of Experimental Psychology, vol. 105, no. 5, pp. 683-703, 1984.

2

E. Gibson, ``Perceptual learning in development: Some basic concepts,'' Ecological Psychology, vol. 12, no. 4, pp. 295-302, 2000.

3

R. Arkin, Behavior-based Robotics.
Cambridge, MA, USA: MIT Press, 1998.
ISBN:0262011654.

4

R. Murphy, ``Case studies of applying Gibson's ecological approach to mobile robots,'' IEEE Transactions on Systems, Man, and Cybernetics, vol. 29, no. 1, pp. 105-111, 1999.

5

Y. Aloimonos, C. Fermüller, and A. Rosenfield, ``Seeing and understanding: Representing the visual world,'' ACM Computing Surveys, vol. 27, no. 3, pp. 307-309, 1995.

6

M. Lungarella, G. Metta, R. Pfeifer, and G. Sandini, ``Developmental robotics: a survey,'' Connection Science, vol. 15, no. 4, pp. 151-190, 2003.

7

P. Fitzpatrick, G. Metta, L. Natale, A. Rao, and G. Sandini, ``Learning about objects through action -initial steps towards artificial cognition,'' in Proceedings of the 2003 IEEE International Conference on Robotics and Automation, ICRA, pp. 3140-3145, 2003.

8

K. MacDorman, ``Responding to affordances: Learning and projecting a sensorimotor mapping,'' in Proc. of 2000 IEEE Int. Conf. on Robotics and Automation, (San Francisco, California, USA), pp. 3253-3259, 2000.

9

A. Stoytchev, ``Toward learning the binding affordances of objects: A behavior-grounded approach,'' in In Proceedings of AAAI Symposium on Developmental Robotics, pp. 21-23, March 2005.

10

A. Stoytchev, ``Behavior-grounded representation of tool affordances,'' in In Proceedings of IEEE International Conference on Robotics and Automation (ICRA), (Barcelona, Spain), pp. 18-22, April 2005.

11

I. Cos-Aguilera, L. Canamero, and G. Hayes, ``Motivation-driven learning of object affordances: First experiments using a simulated Khepera robot,'' in In Proceedings of the 9th International Conference in Cognitive Modelling (ICCM'03), (Bamberg, Germany), 4 2003.

12

G. Dorffner, J. Irran, F. Kintzler, and P. Poelz, ``Robotic learning architecture that can be taught by manually putting the robot to action sequences,'' tech. rep., The Austrian Research Institute for Artificial Intelligence (OFAI), 2005.
MACS Project Deliverable 5.3.1, draft, version 1.

13

G. Fritz, L. Paletta, M. Kumar, G. Dorffner, R. Breithaupt, and R. Erich, ``Visual learning of affordance based cues,'' in From animals to animats 9: Proceedings of the Ninth International Conference on Simulation of Adaptive Behaviour (SAB) (S. Nolfi, G. Baldassarre, R. Calabretta, J. Hallam, D. Marocco, J.-A. Meyer, and D. Parisi, eds.), LNAI. Volume 4095., (Roma, Italy), pp. 52-64, Springer-Verlag, Berlin, 25-29 September 2006.
in press.

14

E. Ugur, M. R. Dogar, O. Soysal, M. Çakmak, and E. Sahin, ``MACSim: Physics-based simulation of the KURT3D robot platform for studying affordances,'' 2006.
MACS Project Deliverable 1.2.1, version 1.

15

K. Kira and L. A. Rendell, ``A practical approach to feature selection,'' in ML92: Proceedings of the ninth international workshop on Machine learning, (San Francisco, CA, USA), pp. 249-256, Morgan Kaufmann Publishers Inc., 1992.

16

I. Kononenko, ``Estimating attributes: analysis and extensions of RELIEF,'' in ECML-94: Proceedings of the European conference on machine learning on Machine Learning, (Secaucus, NJ, USA), pp. 171-182, Springer-Verlag New York, Inc., 1994.

17

V. N. Vapnik, Statistical Learning Theory.
Wiley-Interscience, September 1998.

Footnotes

...BR

This work was partially funded by the European Commission under the MACS project (FP6-004381).

... finder

http://www.ais.fraunhofer.de/ARC/kurt3D/

... Engine)

http://ode.org/

... (SVMs)

The LibSVM software that is used in this study, is available at http://www.csie.ntu.edu.tw/~cjlin/libsvm