Oliver Strimpel

The unifying theme of the Smart Machines gallery is to demonstrate how machines do things that have hitherto been the province of intelligent human activity. We were determined to convey to our visitors the tremendous sophistication of the human mind and body, as well as some of the difficulties scientists face in their attempts to replicate even the simplest of human activities. The combination of A. I. and robotics was straightforward enough: we wanted to demonstrate both the mental capabilities and the physical dexterity of today's machines. This article attempts to explain how the various live exhibits selected for Smart Machines exemplify past and present trends in A. I. and robotics.

The exhibit is grouped into six sections: language understanding, knowledge- based systems, game-playing, robot sensing, mobile robots and robot arms. The historical time-line and robot theater are described in the next article.

Language Understanding

One of the major conclusions of A.I. research during the 1970s was that knowledge and language could not be clearly separated. The early attempts to understand or translate language on a word-by-word basis failed. However, research has continued along several lines, and progress has resulted in commercially successful products.

A grammar correction system from Houghton-Mifflin shows how much a computer can do without any knowledge of the meanings of words. Visitors can watch the grammar checker find mistakes and correct them automatically.

Unlike a grammar checker, parsing is just the first stage of a program that actually tries to understand the meaning of a sentence. In a natural language interface program, the knowledge resides in the database. However, the questions stated in English must be translated into a machine language query to the database. Our exhibit features Datatalker, a natural language interface from Natural Language, Inc. It asks visitors to type in information about themselves, which it stores. It then invites questions in plain English about previous visitors. The program's task is eased because it expects a question about something in its database. After parsing a visitor's question stated in English, the program tries to extract the sentence's meaning, and, if appropriate, converts it into instructions to search through its database for information that will answer the question. The result of the search is translated back into an English reply. Other parts of the program keep track of the dialog, deciding when responses are adequate.

To go beyond a simple question and answer conversation, computers need a much wider and deeper knowledge. The exhibit addresses this enormous problem by demonstrating some of complexities of building a real computer like HAL in the film 2001: A Space Odyssey.

Two exhibits are conversational programs that pretend to know more than they do. ELIZA, the classic computer psychotherapist program written by Joseph Weizenbaum in 1966, takes key words from the visitor's typed-in text and uses them to trigger stock questions. It also repeats the user's words, turning statements into questions. ELIZA exploits its role as a non-directive therapist to justify its extreme passivity. In contrast, RACTER converses volubly with the visitor on many arcane topics. Like ELIZA, it has no model of the world, but responds to key words in the input text by concocting sentences based on standard forms. It attempts to skirt around its lack of understanding by making a virtue out of being zany. These programs are not presented as A.I., but as illustrations of the limitations of approaches that use words without knowledge.

Knowledge-Based Systems

The greatest number of useful applications in the field of A.I. have emerged from rule-based expert systems. Several hundred expert systems perform tasks ranging from diagnosing failures on gas turbines to suggesting which pesticides to use on a particular crop. In general, a Museum should exhibit genuine examples of its subject matter.

However, expert systems are tools aimed at the technical user and would be totally incomprehensible to the majority of our visitors. As a compromise, we included one "real" expert system, somewhat modified for the Museum by its author, Randy Miller. The system is Quick Medical Reference (QMR), a medical diagnosis system that contains descriptions of nearly 600 diseases. Visitors can browse through the system, using it like an electronic textbook indexed either by disease or by symptom. Alternatively, visitors can retreive a patient's case, make QMR diagnose it, and compare QMR's hypothesis with one of their own.

We assembled several highly instructive and entertaining "nonreal" rule-based systems to demonstrate the capabilities and internal workings of expert systems. In the Haymarket exhibit, visitors haggle with up to three different rule-based storekeepers to buy a large box of strawberries. The simplest, Noah Budge, has only 8 rules and never budges on his price. Eventually, he will kick you out of the store if you don't give him what he's asking for. Visitors can choose Ho Nin with 30 rules and

Nora Logical, the sophisticated storekeeper with over 100 rules. Another rule-based demonstration is a wine-advisor. This proceeds via a two-way spoken conversation. Visitors are asked questions about the type of food planned for the meal and what their tastes are in general. They respond by speaking into a microphone. After up to 10 questions, the computer makes a specific recommendation.

Several rule-based systems dealing with the arts are also on display, including a musical score follower (right) and a drawing expert (below). The goal is to demonstrate the application of rule-based programming techniques in non- technical domains. A computer composition system by Charles Ames generates rock and jazz pieces, which it performs through a Kurzweil 250 synthe- sizer. After selecting a musical style and a model, such as the twelve bar blues, the program selects instruments and then com- poses the rhythm, assigning each note a duration that depends in part on whether it is a basic, ornamental or cadence note. Finally, pitches are selected according to a set of about 20 rules. The rules make the notes conform to the harmony, create a melody and avoid repetition. The music is surprisingly convincing.

In contrast to all the systems described above, which represent knowledge as sets of rules, TALE-SPIN is based on scripts. This program came from the work of Roger Schank's group at Yale on language understanding. TALE-SPIN is a program that generates stories with a simple "point" somewhat reminiscent of the simpler Aesop's fables. The program simulates a world of char- acters who do things because they have problems to solve. These consist of fulfilling simple goals, such as satisfying hunger or thirst. Visitors select a main character -- Joe Bear, Irving Bird or Lucy Lamb -- and also determine the goals and character traits of the players. The program has a model of its characters and ensures that their behavior is rational. For example, if Joe Bear is thirsty and sees a river, he will try to get to the river.

In addition to rules, knowledge can be represented as frames, semantic nets and scripts. These are illustrated by panels in the exhibit.

Game-Playing

In addition to being fun, computer games are a valuable testing ground of ways to search through enormous numbers of alternative solutions to a given problem. Typically, when people play a game, they rely on knowledge of the opponent's ability and on an understanding of what it takes to win. Machines, on the other hand, rely on searching many possible moves to determine the best outcome. Research efforts have concen- trated on optimizing the search for moves in chess. One approach is to perform the search in the proper order so that unpromising avenues can be elimi- nated early on. Another approach seeks to give the com- puter knowledge about chess, increasing its ability to "size-up" a given position. Hans Berliner and his colleagues at Carnegie-Mellon University used both approaches to build the world's strongest computer chess player. Their program, called Hitech, has custom hardware to generate and evaluate up to 200,000 moves a second. This enables it to search about 11 half-moves ahead while playing in a tournament. In addition, Hitech's board knowledge is equivalent to a search of a further three half-moves.

Visitors can also play tic-tac-toe and five-in-a-row and choose the computer's strategy to be one of look-ahead search, voting or random. The program offers graphics that give an "X-ray" view of the program's deliberations.

A checker player by David Slate can beat all but the most serious players. Finally, in the game "How the West Was Won," the computer plays two roles: opponent and tutor. This is a numbers game, designed to help children gain familiarity with arithmetic. The computer tutor analyzes one's moves and suggests possible improvements. It never scolds or repeats itself and lets the player discover the game for him or herself. This coach was developed as a robust, friendly and intelligent tutor that could work well in the home and classroom.

Robot Sensing

Giving robots sensory capabilities is an important part of the effort to endow robots with intelligence. A smart robot must find its way independently and cope with the unexpected. It can only begin to do this if it can sense the distance to any surrounding obstacles, feel if it is touching something, or analyze pictures taken with an onboard camera. Many of the historic robots acquired by the Museum and on display in the exhibit's Smart Machines Theater were built as experiments, allowing researchers to explore how a robot can gather and make good use of sensory data.

The Museum visitor can experiment with four robot senses: vision, hearing, touch and sonar. Human vision is so sophisticated that we hardly appreciate its complexity. For example, just consider how we can instantly recognize everyday objects, such as a tree or cat, even though no two examples look alike in detail. Our vision relies on a great deal of knowledge about the world and about what we expect to see. By contrast, machines rely mainly on the details of the actual image, analyzing it first to find edges, and identifying objects by their outlines. This ap- proach makes machines better at matching complicated abstract patterns, such as fingerprints. Part of a fingerprint recognition system used by police departments all over the world is on display. Visitors try to match a fingerprint on the screen with one of several prints displayed on the wall from famous criminals. The computer then shows how it would make the match, using the points where ridges start or fork to classify the pattern accurately.

Speech recognition systems can give computers a reasonable sense of hearing, particularly if the machine has been trained by the speaker. Visitors can use several systems, including one that can be trained to respond to the visitor's voice. Even after training, computer speech recognition is limited to a few thou- sand words at most and generally requires the speaker to pause briefly between each word. In both speech recognition and vision, computers have yet to match the ability of a two-year old child.

A sense of touch is needed by a robot hand when it tries to grasp a delicate object. A pressure sensitive pad mounted on the robot gripper can gauge the amount of pressure being applied. Visitors see the pressure of their fingers on a pad displayed as an array of colors on a screen.

Finally, visitors can try out a sense that humans do not have - sonar. Robots use sonar to gauge the distance to surrounding walls and obstacles. The sensor emits pulses of extremely high-pitched sound, which reflect off an object and are picked up by the detector. The sound's round trip travel time indicates the distance to the object. In the exhibit, a ceiling mounted sensor measures a visitor's height by bouncing a signal off the top of the head.

Mobile Robots

In addition to its sensing ability, an intelligent, independent robot must have a suitable drive system and should be able to form and achieve goals. All the mobile robots on display in the exhibit are equipped with a drive system. Most have some form of sensing, but only Shakey seriously attempted the last and hardest requirement of forming plans and reasoning.

A mobile robot from Real World Interfaces roams around a cage, using sonar to sense and map the walls and obstacles. Visitors can try to override the robot's good sense by controlling its movement with a joystick, but it will never let itself collide with a wall. In addition, about 25 robot toys are on display and can be tried out by visitors. Most have wheels, but several can walk; some have bump sensors, or respond to claps or squeezing.

An application mobile robots have already found is that of night watchman. The gallery's Sentry robot by Denning Mobile Robotics can carry TV cameras, infrared sensors and microphones to detect an intruder. The information it collects is radioed to a security office. Microwave beacons supplement the Sentry's onboard sonar, enabling it to patrol a path hundreds of feet long for hours on end without ever losing an exact knowledge of its position. In the exhibit, the Sentry patrols a short path, avoiding obstacles in its way. Its TV camera relays signals to another robot, the Hubot, whose onboard TV monitor displays the picture.

Robot Arms

Robot arms and hands attempt to replicate aspects of human manual dexterity. Arms are by far the most common type of robot. They perform a wide range of industrial tasks, from the tiny movements for assembling a wristwatch to the large powerful movements required to stack heavy cartons. In the exhibit, the real industrial arms are shown on video, and smaller, educational arms are operated by visitors.

Two robot hands are on display: the five-fingered Tomovic hand attached to the tentacle arm pictured on the front cover, and a three-fingered soft gripper from Shigeo Hirose at the Tokyo Institute of Technology.

An ingenious way to achieve responsive compliance was invented at the Draper Laboratories. Their system uses an arrangement of springs that greatly eases tasks such as putting a peg into a tightly fitting hole. With a stiff wrist, a robot would jam the peg and only make it worse by pushing harder. With the compliant wrist, however, the peg finds its way into the hole smoothly. Visitors can use a compliant wrist to try this out for themselves.

A major thrust of industrial development is to tighten the link between the design and manufacture of a product. Using a computer-aided design system, an industrial designer can create a product and then send instructions for making that product directly to a numerically controlled tool or to a robot. Visitors can experiment with this process by designing a log cabin made of lincoln logs. When the design is complete, the cabin is constructed automatically by a pair of simulated robots on a screen. Real robots would need to be guided by a vision system to ensure that the logs were positioned accurately. This is demonstrated in an adjacent display in which a vision system guides a robot arm that assembles a toy boat from its parts. Both these displays were provided by the University of Lowell's Center for Productivity Enhancement.

The Future

The exhibit can be readily updated as new items become available. A large industrial arm has already been offered to us by Cincinatti Milacron, and we hope to be able to demonstrate an industrial application. We welcome suggestions from our members and visitors!

Gwen Bell and Leah Hutten

The Smart Machines exhibition has two historical components. A timeline, on display at the entrance to the exhibit, chronicles the major milestones to 1979. The Robot Theatre displays a collection of historic robots through the early 1980s. This article is intended as a synthesis of these two exhibits

Precursors

• 1738
  Jacques de Vaucanson builds a mechanical duck to tour and raise money for the inventor's experiments for creating life artificially. The copper duck quacks, bathes, drinks water, eats grain, digests it, and voids.

• 1818
  The book, Frankenstein, by Mary Shelley, includes the first description of creating a manmade being, who becomes a fearful monster.

• 1900
  At the Paris World's Fair, Torres y Quevedo demonstrates an electromechanical machine that can play selected chess end games .

• 1920
  Karel Capek writes the play R. U. R. (Ross='s Universal Robots) in which robots are produced by an Englishman named Rossum. The name, Rossum, is derived from the Czech word for reason, while robot is a Czech word for worker. The popularity of the play led to the widespread adoption of the word robot.

• 1942
  Isaac Asimov publishes "Runaround" in the March issue of Astounding, in which he introduces the Three Laws of Robotics.

  This is the first known use of the term "robotics."

• 1943
  Warren McCulloch and Walter Pitts propose that the behavior of the brain can be treated as a network of neurons that behave like on-off switches.

• 1947
  Only a year after the completion of ENIAC, the first electronic computer, Arthur Samuel proposes to build a computer to play checkers.

• 1948
  Norbert Wiener coins the term cybernetics, a philosophical perspective for describing interacting systems in terms of exchange of information.

• 1949
  The Debate Begins: Can Machines Think?

  On June 9, at Manchester University's Lister Oration, British brain surgeon Sir Geoffrey Jefferson states, "Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain that is, not only write it but know that it had written it. No mechanism could feel (and not merely artificially signal, am easy contrivance) pleasure at its successes, grief when its valves fuse, be warmed by flattery, be made miserable by its mistakes, be charmed by sex, be angry or miserable when it cannot get what it wants."

  On June 11, The London Times quotes the mathematician Alan Turing, "I do not see why it (the machine) should not enter any one of the fields normally covered by the human intellect, and eventually compete on equal terms. I do not think you cam even draw the line about sonnets, though the comparison is perhaps a little bit unfair because a sonnet written by a machine will be better appreciated by another machine."

  In New York, Claude Shannon's paper to the Institute of Radio Engineers proposes two computer chess strategies that are still in use. The first is to look at all the choices up to a fixed depth and the second is to look at a selected few to greater depth.

• 1951
  Turing creates a standard test to answer: "Can machines think?" If a computer, on the basis of written replies to questions, could not be distinguished from a human respondent, then it must be "thinking,"

• 1954
  George C. Devol, Jr., applies for the first US patent for an industrial robot. He calls it "unimation" for short.

• 1956
  John McCarthy of Dartmouth convenesthe Dartmouth Summer Research Project on Artificial Intelligence, marking the birth of the field.

  Herbert Simon, Allen Newell, and J.C. Shaw write "Logic Theorist," one of the earliest programs to investigate the use of heuristics in problem solving.

• 1957
  John McCarthy and Marvin Minsky found the first artificial intelligence laboratory at M.I.T.

  Simon, Newell and Shaw write the pioneering, "General Problem Solver." It is the first program that solves a problem that it hadn't been specially programmed to solve.

• 1958
  Simon, Newell and Shaw design and use the first list processing program, IPL-V.

• 1959
  McCarthy creates LISP. Unlike other current programming languages, LISP is designed to work with English words and phrases. A key feature is that the data and programs are simply lists in parentheses, allowing a program to treat another program - or itself - as data. This characteristic greatly eases the kind of programming that attempts to model human thought.

  Frank Rosenblatt invents an ingenious evidence-weighing machine called a "Perception." It is supposed to recognize patterns by their parts without regard to their relationships.

Joe Engelberger, the entrepreneur, works tirelessly to get Joseph Devol's ideas for industrial robots into use. Engelberger eventually eams the title "Father of Robotics. "

• 1961
  James Slagle writes a Symbolic Automatic Integrator (SAINT) to solve elementary symbolic integration problems at the level of a good college freshman.

  SAD-SAM (Syntactic Appraiser and Diagrammer Semantic Analyzing Machine) is programmed by Robert Lindsay at Car negie Institute of Technology. The program accepts English sentences about kinship relations, builds a data base and answers questions about the facts it has stored.

  SAD-SAM INPUT: John is Mary's son. SAD-SAM OUTPUT: Mary's brother is John's uncle; Mary's mother is John's grandmother, etc.

• 1962
  Engelberger founds Unimation, the first industrial robot company. The Unimate Mark II robot welcomes visitors into the Museum's Smart Machines Gallery.

  Samuel's checkers program, which has the ability to learn from its mistakes, plays at the masters level.

• 1963
  McCarthy leaves M.I.T, and founds Stanford University's artificial intelligence laboratory.

  At Rancho Los Amigos Hospital an orthotic arm is designed to aid a paralyzed person. Stanford University modifies the Rancho Arm to be controlled by a computer. Photo Dan McCoy, Rainbow. On loan from Stanford University

• 1965
  A five fingered aluminum prosthetic hand is developed by Rajko Tomovic at The University of Belgrade.

  The PDP-6 becomes the workhorse machine for the artificial intelligence community. The PDP-6's architecture is particularly suited for running LISP programs.

  Edward Feigenbaum and Bruce Buchanan conceptualize expert systems and start the Dendral project.

  Hubert Dreyfus' paper "Alchemy and Artificial Intelligence," is published by The Rand Corporation. His assertation that "Even though machines can perform intelligent tasks, the evidence against their ever becoming able to be really, humanly intelligent, is overwhelming," leads to debates and research that continue into the 1980s.

  Scientists at Johns Hopkins University create the Beast (Mod II) as an experiment to replicate animal behavior in a robot. When it gets "hungry" (low batteries), it uses sonar and a photocell array to find "food" (wall power outlets). Pressure-sensitive switches perform the fine guiding of its prongs into the wall socket.

  Victor Scheinman and Larry Leifer build the "Orm," Norwegian for snake. This robot arm moves by selectively inflating groups of its 28 air sacks sandwiched between seven metal disks. Its design is later abandoned because its movements could not be repeated accurately.

  The Stanford Cart is built at the A. I. Lab to simulate a remotely controlled Moon rover.

• 1966
  The program ELIZA, written by Joe Weizenbaum at M. I. T., tries to assume the role of a nondirective therapist. It turns sentences into questions and responds to key words about feelings and family.

  Richard Greenblatt's MacHac is the first machine to achieve a Class C rating in the National Chess Association (approaching the level of a serious weekend amateur player).

• 1967
  Television cameras controlled by a remote computer are added to the Stanford Cart, permitting it to follow a white line on a road.

• 1968
  Engelberger travels to Japan and grants Kawasaki the right to build Unimates in exchange for royalties. These are the first robots built in Japan.

  Bruce Buchanan and Edward Feigenbaum, working with Chemist card Nobel Laureate Joshua Lederberg, complete DENDRAL, an expert system for generating explanatory hypotheses in organic chemistry.

  Marvin Minsky constructs the 12 jointed Tentacle Arm, which can reach around obstacles. A PDP-6 computer is used for control and hydraulic fluids for power.

  Seymour Papert writes "The Artificial Intelligence of Hubert L. Dreyfus: A Budget of Falacies." In it, he states, "It is cowardice to ... assure us that the computer is barred by its finite number of states from encroaching further into areas of activity ... (regarded) as 'uniquely human'."

• 1969
  Shakey, the first integrated robot system equipped with a TV camera and other sensors, slowly roams through the roorns of The Stanford Research Institute, guided by the remote radio control of an SDS-940 computer.

  The Original Stanford Arm, the first successful electrically powered computer-controlled robot arm, is constructed by Victor Scheinmam. It is used to develop industrial assembly techniques for robots.

• 1970
  200 people attend the first meeting of the International Joint Conference on Artificial Intelligence (IJCAI).

  Terry Winograd integrates natural language understanding and knowledge about a world of table top blocks in SHRDLU, written for his doctoral thesis at M.I.T.

• 1971
  France installs its first industrial robot, a Unimate, at Renault's R-5 plant to build LeCar.

  Stanford Research Institute gives Shakey the ability to reason about its actions. Shakey radios information from its sonar and bump sensors to a room-sized computer (DEC PDP-10 and PDP-15), which sends back commands to make Shakey move. The computer spends about half am hour to move Shakey one meter.

  DARPA funds a $15 million, fiveyear research program to achieve a breakthrough in speech understanding.

• 1972
  At the University of Aix-Marseille, Alain Colmerauer develops the use of formal logic as a programming language, PROLOG.

  LUNAR, a natural language information retrieval system, is completed by Woods, Kaplan, and NashWebber at Bolt, Beranek and Newman. LUNAR helps geologists access, compare and evaluate chemical-analysis data on moon rock and soil composition from the Apollo 11 mission.

• 1973
  Yorick Wilks writes the first acceptable language translation program, which produces respectable French from small English paragraphs.

• 1974
  The first commercially available mini-computer controlled robot, T3, is produced by Cincinnati Milacron.

  The first World Computer Chess Tournament is held.

  CONS, the first computer built to optimize LISP, is completed by Tom Knight at M.I.T.'s A. I. Lab. It is the precursor of CADR and the commercial machines built at LMI and Symbolics.

• 1975
  MARGIE (Meaning Analysis, Response Generation, and Inference in English) is developed by Roger Schank and his students at the Stanford A. I. Laboratory.

  Minsky develops the concept of frames as a convenient way to represent specific objects or concepts. Each frame consists of a name and a series of slots that describe the object's or concept's attributes.

  Unimation has its first profitable year.

• 1976
  The DARPA speech goals are met by the HEARSAY speech program developed at Carnegie-Mellon University under the direction of Raj Reddy. It beats DARPA's goal of understanding 90% of ordinary continuous speech using a vocabulary of 1000 words.

• 1977
  Hans Moravec equips the Stanford Cart with stereo vision. A television camera that moves along a rail takes pictures of a given scene from several different angles, enabling the Cart to find the distance to obstacles in its path.

  In the USA, Robert McGhee develops a hexapod walking machine controlled by a digital computer. In the USSR, scientists develop a hexapod walker controlled by a hybrid (analog and digital) computer.

  EMYCIN developed by William Von Melle, Edward Shortliffe, Bruce Buchanan, and Edward Feigenbaum is the first expert system "shell." A shell is a program that provides the framework for developing an expert system. The user supplies his own rules to build an expert system in the subject of his choice.

  The programs SAM (Script Applier Mechanism) and PAM (Plan Applier Mechanism) are developed by Roger Schank, Robert Abelson and their students at Yale University. SAM and PAM demonstrate the understanding of stories by using scripts and plans.

  The Jet Propulsion Laboratory builds two Rover prototypes designed to explore Mars. To stay upright, the Hardware Prototype has caterpillar tracks mounted on flexible legs.

  The Software Prototype has both sensing ability and intelligence.

• 1978
  The Mars Rover project is cancelled because NASA opts for a manned space program.

  GM unveils its production line, which uses a programmable universal machine for assembly (PUMA) system based on the Scheinman arm.

  Hans Berliner's backgammon program wins the world championship.

  Consight-I, by Steven Holland, Lothar Rossol and Mitchel Ward, is able to identify and sort randomly oriented parts on a moving factory conveyor belt. When the parts move under two converging light beams, the beams are split in two. This pattern is detected by a computer connnected to a television camera. Consight-I consists of a Vicarm robot arm controlled by a PDP 11/45 computer. Commerical versions use an arm by Cincinatti Milacron.

• 1979
  After seven years of research, Moravec's refined Stanford Cart successfully traverses, without human intervention, a room strewn with chairs

• 1981
  Avatar represents a breed of personal home robots that evolved following the microcomputer revolution in the early 1980s. This robot can move without bumping into things, talk, and handle objects with its arm.

• 1981
  The first direct drive (DD) arm by Takeo Kanade served as the prototype for DD arms used in industry today. The electric motors housed inside the joints eliminate the need for chains or tendons used in earlier robots. DD arms are fast and accurate because they minimize friction and backlash.

• 1981
  One of three of Shigeo Hirose's robots at the museum, the quadruped can perform a complicated task such as "feeling its way" up stairs of varying heights. It has contact sensors on the sides and bottom of its feet. When these are touched, the quadruped responds with animal-like reflexes. Each leg contains an elegant mechanical device that translates small motor movements inside the body into larger movements of the legs.

• 1983
  

  Odex is the first commercially available walking robot. It can work in dangerous places inaccessible to vehicles with wheels. These include radioactive zones in nuclear power stations, military battlefields, and underground mines. Its legs can also serve as arms for lifting and moving objects.

  in an e-mail to Ed Thelen, May 23,2007, David Buckley wrote: "... It has the wrong photo for the Odex walker, instead it has Avatar featured in Robotics Age probably Jan/Feb 82. Avatar has wheels not legs! I remember it well, my robot Quester was featured in Robotics Age Jan/Feb 82 as part of the same competition which Avatar won. www.davidbuckley.net/DB/QuesterRoboticsAge.htm

  Odex is the first commercially available walking robot. It can work in dangerous places inaccessible to vehicles with wheels. These include radioactive zones in nuclear power stations, military battlefields, and underground mines. Its legs can also serve as arms for lifting and moving objects.

• 1985
  Underwater rovers explore the ocean depths under remote control. The famous Titanic wreck was explored by a large rover called Argo, but most underwater rovers are used for more routine inspection tasks. The Sea Rover, designed by Christopher Nicholson, can dive to depths of up to 120 meters and travel at 1.5 knots while relaying color video pictures from under the sea.

BIPER-3, designed by Hirofumi Miura and Isao Shimoyama in 1981, was the first legged machine to balance itself dynamically. Like a person, its gait relies on its own forward momentum. It has stilt-like legs and uses its hips to pick up its feet. This gives the machine a pronounced shuffling gait like Charlie Chaplin's stiffkneed walk. BIPER-3 can walk forward, backward, or sideways. On loan from the University of Tokyo, Tokyo, Japan