The problem of instructing a robot is similar to that of instructing a child, but much more difficult, because the robot has no linguistic background to understand declarations and interpret instructions for specific actions. Some 30 years ago, Giuglio Pasqualigo, in a communication sent to the Brazilian Academy of Science, made a very curious inquiry: Could a population of intelligent robots (or “cybes,” according to his designation) exhibit or simulate the behavior of a human society? The answer was negative because an artificial brain lacks at least 20 basic thinking principles, such as the notion of number (and the very possibility to elaborate general concepts starting from perceptions), that of infinity, the basic axioms of arithmetic (according to Peano) and geometry (due to Hilbert), the logical mechanisms of deduction, and those, much more complex, of induction; besides, an artificial brain lacks the capacity to feel emotions and to manifest will or intention. Nevertheless, analyzing the problems related with the instruction of an artificial mind (robot), one can identify the necessary steps to progressively create and enlarge the missing background for intelligent and concise communication through natural language.

In preceding work (e.g., Winograd’s SHRDLU program [1], the HEARSAY-II [2], HARPY [3], and Miller and Johnson-Laird [4]), the principal objective was the understanding of the speech uttered in natural language. The present work aims at the general and more ambitious question of designing systems capable of building new procedures out of already known ones. This is approached by using interactive features somehow similar to the interactive theorem provers and much more flexible and elaborate than simple “macro” instructions. The basic experimental system is comprised of the following:

• a grammar and a parser (consisting of a preparser, a main parser, and a postparser);

• the translation from parsed English to operator language, providing a semantic interpretation for the new primitive LISP-atoms that can occur: operators (e.g., ADD), actors (e.g., DEFAULT-NUMBER), and data (e.g., TOP);

• the execution of procedures, including an interpretation and full analysis of the execution, with references to past steps; and

• a learning procedure, enabling the change in behavior, without any restriction imposed on the English as source language; hence, it allows ambiguities, which are resolved by asking questions.

The first program was limited to the teaching of addition and subtraction by columns. Various improvements appeared necessary before extending the system to other forms of knowledge representation; but the results seem valid and encouraging, notwithstanding the limitation to a very restricted domain of the procedures involved.

Further developments the authors hope to accomplish in the future will certainly entail (1) the necessity to use a very large memory, (2) the adoption of specialized techniques for classification and retrieval of abstract concepts, and (3) intensive use of parallel programming; then they may be able to imitate the extraordinary capacity of the human brain.