Obstacle Avoidance and Navigation in the Real World by a Seeing Robot Rover, Hans Moravec, 1980
<-- Previous  Index -->

Appendix 12: Connections

Philosophy

Artificial Intelligence researchers are like the blind men who went to see the elephant. Having of necessity experienced only tiny portions of the situation, each comes to a different conclusion about the nature of the whole.

One group suggests that the construction of an intelligent machine is very like mathematics, and finding the “theorems” of intelligence will involve clever representations, transformation rules and long lemmas, gotten at mostly by thinking hard.

Another school feels AI is like theoretical physics, and the solution involves finding the universal “laws of intelligence”, by means of theories guided by experiment.

Yet another sees AI as a little like biology, the idea being to explain intrinsically complicated natural mechanisms as simply as possible.

A fourth treats AI as a problem in psychological introspection, transferring rules of conscious thinking into mechanical form.

A fifth feels the problem is one of engineering, with an artificial intelligence being just another big machine, to be built subsystem by subsystem, by rule of thumb and experience based intuition.

People's points of view change with experience and mood, and most of us have found ourselves espousing different approaches at different times.

The AI effort has a specific and lofty goal, the matching of human performance in intellectual and other tasks by machines. Because the overall goal is still far from accomplished, many of us suffer from doubts about our progress. These often expresses themselves in the feeling that much of our field is somehow not “scientific”. Depending on our mood, this transforms to “not like mathematics”, or "not like physics" etc.. And of course its easy to find many projects that fail to meet our arbitrary standards, and confirm our suspicions.

The hard sciences are distinguished from many other intellectual pursuits not by the quality of the workers, or even the methods employed, but by the amount of independent verification and refutation practiced. It is the ruthlessness of the evaluation function that separates the useless from the valuable and the capable from the incompetent.

I feel it is too early to commit ourselves to or to excessively condemn any of the various approaches. We ought to judge AI programs on the basis of performance. Whether or not they conform to our theory of the moment as to what constitutes intelligence and how to go about building it, or what is esthetic, we should ask “how well does it work?”.

In other words, I think AI is very like evolution. We should try different modifications and approaches and see which ones prove themselves experimentally.

Since this is in itself a prejudgement, I don't really want to force it on others. But if we suspend disbelief for a few minutes, I can use it to show why roving vehicles are on the direct path to human equivalence. The argument is by analogy with natural evolution.

Locomotion, Vision and Intelligence

Consider that, with few exceptions, the only natural systems with AIish capabilities are large mobile animals. An apparent minimum size for nerve cells explains the complexity limits on small animals like insects. The role of mobility in the development of imaging vision and intelligence is more subtle, yet real. No plants or sessile animals (what few there are) have imaging eyes or complex nervous systems, but there are several independent instances of vision and comparative intelligence in the presence of mobility.

The evolutionary mainstream (as defined by us mainstreamers), fishes through amphibians and reptiles to mammals to us, is one such instance. Imaging eyes and a moderate brain developed roughly simultaneously with a backbone, in motile protofish, sometime in the Paleozoic, about 450 million years ago. Brain size changed little through the slow moving amphibian and reptile stages, then accelerated sharply with the transition to the more mobile mammalian form, about 100 million years ago.

Instance two is the birds, who also have reptilian ancestry, and who's development parallels our own. Though size limited by the dynamics of flying, several bird species can match the intellectual performance of all but the smartest mammals. The battle of wits between farmers and crows is legendary, and well documented. The intuitive number sense of these birds goes to seven, compared to three or four with us (without counting). Hard evidence comes from “reversal learning” experiments. The response giving the reward in a Skinner box is occasionally inverted. Most animals are confused by the switch, and actually take longer than the first time to learn the new state (as if they first had to unlearn the old rules). Primates (monkeys and apes and us) among mammals, and virtually all birds, on the other hand, “catch on” after the first reversal, and react correctly almost instantly on later swaps. Instance three is surprising. Most molluscs are nearly blind, intellectually unimpressive, very slow moving shellfish. Their relatives who opted for mobility, the cephalopods (octopus and squid) provide a dramatic contrast, having speed, good eyes, a large brain, a color display skin, mammal-like behavior, and even manipulators. The similarities to mammals are especially significant because they were independently evolved. Our last shared ancestor was a billion year old bilaterally symmetric pre-worm, with a few neurons. The differences are interesting. The eyes are hemispherical and firmly attached to the surrounding skin, and the light sensitive cells in the retina point outwards, towards the lens. The brain is annular, encircling the esophagus, and is organized into several connected clumps of ganglia, one for each arm. A Cousteau film documents an octopus' response to a “monkey and bananas” problem. A fishbowl sealed with a large cork, and containing a small lobster, is dropped into the water near the animal. The octopus is immediately attracted, seemingly recognizing the food by sight. It spends a while probing the container and attempting to reach the lobster from various angles, unsuccessfully. Then, apparently purposefully, it wraps three or four tentacles around the bowl, and one about the cork, and pulls. The cork comes free and shoots to the surface, and the octopus reaches a free tentacle into the bowl to retrieve the lobster, and eats.

The Point

The point of the preceeding ramble is; moving through the wide world is a demanding task, and encourages development of complex responses in those who undertake it. Moving organisms (and machines) must learn to deal with a wide variety of situations, and have many responses open to them. This variety places a great premium on general techniques, and makes highly specialized methods, which may be optimal for sessile creatures, less valuable. These forces seem to have led to relative intelligence in animals. Perhaps they mark one route to the same goal for machines.

References

\refe{A1}{Arnold, R.D.}{Local Context in Matching Edges for Stereo Vision}{ Proceedings: Image Understanding Workshop, (L. S. Baumann, ed.), Science Applications, Inc., Cambridge, Mass., May 1978, 65-72}

\refe{B1}{Brooks, Rodney A.}{}{personal communication, Stanford University, 1977}

\refe{B2}{Binford, Thomas O.}{}{personal communication, Stanford University, 1975}

\refe{B3}{Brooks, Rodney A.}{An Onboard Microprocessor for The Cart}{unpublished, Stanford University, December 1977}

\refe{B4}{Baker, H. Harlyn}{Three-Dimensional Modelling}{Proceedings of the fifth IJCAI, MIT, Cambridge, Mass., August 1977, 649-655}

\refe{B5}{Baumgart, Bruce G.}{Geometric Modelling for Computer Vision}{ (Ph.D. thesis), Stanford AI lab memo AIM-249, October 1974}

\refe{BC1}{Burr, D.J., and R.T. Chien}{A System for Stereo Computer Vision with Geometric Models}{ Fifth International Joint Conference on Artificial Intelligence, MIT, August 1977, 583}

\refe{D1}{Dijkstra, E.W.}{A Note on Two Problems in Connection with Graphs}{ Numerische Mathematik, 1, 269-271, 1959}

\refe{G1}{Gennery, D.B.}{Object Detection and Measurement Using Stereo Vision}{ Sixth International Joint Conference on Artificial Intelligence, Tokyo, August, 1979, 320-327}

\refe{G2}{Gennery, D.B.}{A Stereo Vision System for an Autonomous Vehicle}{ Proceedings of the 5th IJCAI, MIT, Cambridge, Mass., 1977, 576-580}

\refe{GM1}{Grimson, W.E.L. and D. Marr}{A computer implementation of a theory of human stereo vision}{Proceedings of the April 1979 Image Understanding Workshop, 41-47}

\refe{H1}{Hannah, M. J.}{Computer Matching of Areas in Stereo Images}{Stanford Artificial Intelligence Laboratory, AIM-239, July 1974}

\refe{HNR1}{Hart, P., N.J. Nilsson, and B. Raphael}{A formal basis for the heuristic determination of minimum cost paths}{IEEE Transactions on Systems Science and Cybernetics, SSC-4, 2, July 1968, 100-107}

\refe{J1}{Julesz, B.}{Foundations of Cyclopean Perception}{University of Chicago Press, Chicago, 1971}

\refe{K1}{Knuth, Donald E.}{Tau Epsilon Chi, A System for Technical Text}{ Stanford AI lab memo AIM-317, September 1978}

\refe{L1}{Lowry, Michael}{}{personal communication, Stanford University, 1980}

\refe{LOY1}{Levine, M.D., D.A. O'Handley, and G.M. Yagi}{Computer Determination of Depth Maps}{Computer Graphics and Image Processing 2, 1973, 131-150}

\refe{LW1}{Lozano-P\'erez, Tom\'as and Michael A. Wesley}{An Algorithm for Planning Collision-Free Paths Among Polyhedral Obstacles}{ CACM 22, number 10, October 1979, 560-570}

\refe{M1}{}{Macsyma Reference Manual}{version 9, The Mathlab Group, Laboratory for Computer Science, MIT, Cambridge, Mass. December 1977}

\refe{M2}{Moravec, H.P.}{Visual Mapping by a Robot Rover}{Proceedings of the 6th International Joint Conference on Artificial Intelligence, Tokyo, Japan, August 1979, 589-601}

\refe{M3}{Moravec, H.P.}{Towards Automatic Visual Obstacle Avoidance}{ Proceedings of the 5th IJCAI, MIT, Cambridge, Mass., 1977, p. 584}

\refe{MP1}{Marr, D. and T. Poggio}{Cooperative Computation of Stereo Disparity}{ Science 194, Oct 1976, 283-287}

\refe{MP2}{Marr, D. and T. Poggio}{A Computational Theory of Human Stereo Vision}{ MIT AI lab memo 451}

\refe{MKA1}{Mori, K., M. Kidode, and H. Asada}{An Iterative Prediction and Correction Method for Automatic Stereocomparison}{Computer Graphics and Image Processing 2, 1973, 393-401}

\refe{N1}{Nevatia, R.K.}{Depth Measurement by Motion Stereo}{Computer Graphics and Image Processing, 5, 1976, 203-214}

\refe{N2}{Nitzan, D.}{Stereopsis Error Analysis}{Artificial Intelligence Center Technical Note 71, SRI International, Menlo Park, Ca., September 1972}

\refe{PT1}{Pingle, K.K., and A.J. Thomas}{A Fast, Feature-Driven Stereo Depth Program}{Stanford Artificial Intelligence Laboratory, AIM-248, May 1975}

\refe{Q1}{Quam, L.H.}{Computer Comparison of Pictures}{Stanford Artificial Intelligence Laboratory, AIM-144, May 1971}

\refe{QH1}{Quam, L.H., and M.J. Hannah}{Stanford Automatic Photogrammetry Research}{Stanford Artificial Intelligence Laboratory, AIM-254, December 1974}

\refe{R1}{Reiser, John F.}{SAIL}{Stanford AI lab memo AIM-289, August 1976}

\refe{S1}{}{Microbroadcasting}{Scientific American, Vol. 242 $\#5$, May 1980, pp 84-87}

\refe{S2}{Schmidt, R.A.}{A Study of the Real-Time Control of a Computer Driven Vehicle}{(PhD thesis), Stanford AI Memo AIM-149, August 1971}

\refe{W1}{Westphal, Harald}{}{personal communication, Stanford university March 1979}

\refe{W2}{Weyrauch, Richard W.}{}{personal communication, Stanford University, 1975}

\refe{WG1}{Wright, F.W. and R.E. Gorin}{FAIL}{Stanford AI lab memo AIM-291, October 1976}

\refe{YC1}{Yakimovsky, K. and R. Cunningham}{A System for Extracting Three-Dimensional Measurements from a Stereo Pair of TV Cameras}{Computer Graphics and Image Processing 7, 1978, 195-210}

<-- Previous  Index -->