The term “Embodied AI” is having its moment in the sun right now. For me, that is what I have spent my career working on, and I have repeatedly published articles using the term “embodied intelligence”.
I recently stumbled across a piece that I wrote in May 2011 intended for the proceedings of a conference held in June 2012 in Manchester, England, to celebrate the 100th anniversary of Alan Turing’s birth. There was a series of invited talks and a number of other components. I attended the conference but I cannot find any record of a proceedings having been published for the 17 invited talks. Here is what I had written for it, unchanged except for fixing many typos.
The Case For Embodied Intelligence
Rodney A. Brooks, May 2011
Abstract: In 1948 in Intelligent Machinery Turing made the distinction between embodied and disembodied intelligence. While arguing that building an embodied intelligence would be a “sure” route to produce a thinking machine he rejected it in favor of disembodied intelligence on the grounds of technical practicalities of the era. Modern researchers are now seriously investigating the embodied approach to intelligence and have rediscovered the importance of what Turing called “cultural search” in that same paper.
For me Alan Turing’s 1948 paper Intelligent Machinery was more important than his 1950 paper Computing Machinery and Intelligence.
At the beginning of Intelligent Machinery Turing provided counter arguments to a number of possible objections to the idea that machines could be intelligent. And right at the end he introduced a precursor to the “Imitation Game”, now commonly referred to as the Turing Test, of his 1950 paper. In this earlier version, one human not very good chess player would try to guess whether he was playing against another human not very good chess player, or against an algorithm. [[At the time the opponent person had to be not very good so that it didn’t outshine the then current abilities of mechanical chess playing. Today the opponent person would have to be a world champion to have any chance at not being outshone by the mechanical system!]] Expansion of these bookends became the body of Computing Machinery and Intelligence.
Intelligent Machinery itself was not published until 1970, so many early computer science researchers were unaware of it. I was fortunate to come in contact with it right as I was starting my academic career.
The bulk of the paper gives examples of how simple computational mechanisms could be adaptable, could be taught, and could learn for themselves. The examples and mechanisms Turing used in this exposition were networks of active computational elements. Although he connected them back to the universal machines of his 1936 paper, it is remarkable, in hindsight, how different this abstraction was than the one he had previously introduced, of the central processing element with a tape memory–still the essential model for all modern digital computers. Here, instead, he used a model inspired by brains. One can only wonder how different our technological world might be if Turing had lived to fully develop this set of ideas himself. Others carried on this second tradition, but one must think that perhaps Turing’s intellectual influence might have been stronger as he would have been arguing against the approach that was adopted from his earlier work.
For me, the critical, and new, insights in Intelligent Machinery were two fold.
First, Turing made the distinction between embodied and disembodied intelligence. While arguing that building an embodied intelligence would be a “sure” route to produce a thinking machine he rejected it in favor of disembodied intelligence on the grounds of technical practicalities of the era. Second, he introduced the notion of “cultural search”: that people’s learning largely comes from the culture of other people in which they are immersed.
Modern researchers are now seriously investigating the embodied approach to intelligence and have rediscovered the importance of interaction with people as the basis for intelligence. My own work for the last twenty five years has been based on these two ideas.
Turing justifies the possibility making a thinking machine by “the fact that it is possible to make machinery to imitate any small part of a man”. He uses the implicit idea of his universal computing machines to dismiss the idea that it is necessary to emulate a person at the neural signal level in order to have intelligence, and instead suggests a digital computer, “if produced by present techniques, would be of immense size”, which would control a robot from a distance. That robot would be built by “tak[ing] a man as a whole and to try to replace all parts of him by machine”. In particular he suggests the parts would include “television cameras, microphones, loudspeakers, wheels and `handling servo-mechanisms’ …”. Turing’s description from over sixty years ago, fairly precisely describes what is done today in dozens of research labs around the world with our PR2 robots, or Mekabots, with their brains off board in racks of Linux boxes, or even off in the computing cloud.
Turing further rightfully notes that even in building such a robot “the creature would still have no contact with food, sex, sport, and many other things of interest to the human being”. Nevertheless he suggests that such an approach “is probably the `sure’ way of producing a thinking machine”, before dismissing it as too slow and impractical. He suggests instead that it is more practical, certainly at that time, to “see what can be done with a `brain’ which is more or less without a body”. He suggests the following fields as ripe for exploration by disembodied intelligence:
(i) Various games, e.g., chess, noughts and crosses, bridge, poker
(ii) The learning of languages
(iii) Translations of languages
(iv) Cryptography
(v) Mathematics.
With these suggestions much of the early directions for the field of Artificial Intelligence were set, and certainly the odd numbered of Turing’s suggestions formed a large part of the work in AI during its first decade.
In one paper Turing both distinguished embodied versus disembodied approaches to building intelligent machines, praised the former as more likely to succeed and either set or predicted the disembodied directions that were actually followed for many years.
But later, towards the very end of Intelligent Machinery he comes back to the place of bodies in the world. He distinguishes three kinds of search as ways to build intelligent systems: intellectual search, genetic search, and cultural search. The first is the direction that classical AI went, where programs try to learn and improve their performance. Although he did not suggest that it be mechanized, genetic search has become a thoroughly practical approach to design and optimization. And lastly by cultural search, Turing means the way in which interactions with others contributes to the development of intelligence. This developmental approach, using social robots, has only now become practical in the last fifteen years, and is a rich source of both theoretical and practical learning systems for robots.
It is humbling to read Alan Turing’s papers. He thought of it all. First.