Learning Robot

Abstract:

Can we expect the time will come when a robot can acquire a natural language like a child? Although controversial, we believe it will [1]. The research of a learning robot aims at realizing an artificial brain that exhibits such intelligence that a child exhibits. Born in an unknown environment, a learning robot is expected to acquire skills how to act in the environment, the knowledge how the environment is like, and elementary capabilities of language processing.

Research Perspective

The research will be done in the three stages: skill acquisition, pattern-based thinking, and symbol-based thinking. We use neural and evolutionary computations as key technologies.

Skill Acquisition

Skill acquisition should go ahead of language acquisition. Among basic skills are parrot-like speaking using their own voices, exploratory wandering of an environment, map building using fragmentary data of the environment, arm control through inverse kinematics learning, etc.

As for parrot-like speaking, after making robot's own voice by analyzing voice owner's speech data, we implemented a mimicking mechanism which generates robot's own speech as close to target speech as possible [2]. Here our optimal vector quantizer [3] is used for clustering tasks.

 

Figure 1: Learning robot Robin-3

As for wandering, our experiments showed that a neural version of a classifier system [5] quickly learns the wall following task from scratch (Fig. 2). As for map building, we proposed a learning procedure which generates a moderately succinct wall map from fragmentary sonar data obtained during exploration of the environment [4].

 

Figure 2: Learning example of wall following

Future Work

Using a revised classifier system, we will keep investigating a learning mechanism which can find important concepts through skill acquisition.

Contact: Ryohei Nakano, Email: nakano@cslab.kecl.ntt.co.jp

Bibliography

1

Nakano, R.: AI Renaissance (in Japanese), in AI Capriccio, NTT Press, pp. 15-27 (1992).

2

Nakano, R., Ueda, N., Saito, K. and Yamada, T.: Parrot-like speaking using optimal vector quantization, Proc. of International Conference on Neural Networks(ICNN '95), pp. 2871-2875 (1995).

3

Ueda, N. and Nakano, R.: A new competitive learning approach based on an equidistortion principle for designing optimal vector quantizers, Neural Networks, Vol. 7, No. 8, pp. 1211-1227 (1994).

4

Nakano, R., Ueda, N., Saito, K. and Takahashi, M.: Wall map building from fragmentary sonar data, Proc. of International Workshop Robolearn'96, pp. 84-89 (1996).

5

Holland,J.H.: Escaping brittleness: the possibilities of general-purpose learning algorithms applied to parallel rule-based systems, Machine Learning Vol. 2, pp. 593-623 (1986).