your bar-mitzvah, then you can be pretty sure that every small flat box contains another pen-and-pencil set. If you don't come to any such quick conclusion, you might examine the mysterious package more closely, look at the tag or postmark, lift it to feel its weight, shake it to see if it rattles or sloshes, and you usually will have a pretty good chance of perceiving what's inside without ever seeing it. Similarly, it is unfair to expect a computer to recognize real objects unless it first knows something about the expected characteristics of the objects, such as their size, shape, color, and the normal physical relationships among them. Many of the past computer-vision projects tried to "simplify" their tasks by aiming their TV cameras only at artificial objects like boxes and wedges, which had straight-line edges and clear mathematical descriptions. Unfortunately, such objects have objects have few expected sizes or shapes, no normal physical relationships, and rarely any context to guide the recognition process. Because of this, paradoxically, the attempt to simplify may actually have made the recognition problem more difficult. Current research is turning to more natural pictures that may incorporate curved objects and complex surroundings. "Scene understanding systems" are now being built that coordinate the use of several kinds and sources of knowledge in order to solve complex problems. For example, knowledge of illumination, distance measurements, color, spatial relationships, and physical constraints, can all contribute to the accuracy of the interpretation of visual data. Robots I am not going to define the word "robot" here, because of the wide range of interpretations it has. The following examples indicate the general kinds of devices that we shall consider. Without getting enmeshed in the technical details of how they work, let's look at what some of these systems were capable of doing a few years ago. At Hitachi Central Research Laboratory a TV camera was aimed at an engineering plan drawing of a structure built out of various-shaped blocks. A second camera looked at the blocks themselves, which were spread out on a table. The computer "understood" the drawing, reached towards the blocks with its arm, and built the structure. At MIT, the camera was not shown a plan; instead, it was shown an example of the actual structure desired. The computer figured out how the structure could be constructed, and then built an exact copy. At Stanford University, the hand obeyed spoken directions. For example, if someone said into the microphone, "Pick up the small block on the left," that is precisely what the arm would do. At the University of Edinburgh, a jumble of parts for two wooden toys was placed on the movable table near the camera. "Freddy," the Edinburgh hand-eye-table robot system, carefully spread out the parts so that it could see each one clearly, and then, with the help of a vise-like work station at one corner of the table, assembled first the toy car and then the toy boat. At SRI, Shakey the mobile robot was told to "PUSH THE BOX OFF THE PLATFORM." Shakey had no arm, and realized that he could not reach the box unless he was on the platform with it. He looked around, found a ramp, pushed the ramp up against the platform, rolled up the ramp, and then pushed the box onto the floor. Recently, robot researchers have been concentrating their efforts upon specific technical problems that must be solved in order to create more powerful robot systems. Major developments coming out of current work include: (1) new hardware technology that is leading to more reliable and less expensive sensors, effectors, and computers; (2) new software technology, in the form of high-level programming [image] tools and studies of how to structure the large knowledge bases that are essential for any intelligent system; and (3) prototypes of simple robot systems that can at least begin to perform truly practical tasks. For example: At Stanford the hand-eye system that used to stack toy blocks can now assemble a real water pump. At SRI a computer-controlled Unimate industrial manipulator arm with touch and force sensors can feel its way as it packs assembled pumps into a case. At MIT programs are under development to enable a computer to inspect and repair circuit boards for use in computers, TV sets, and other electronic equipment. Applications As computers become less expensive and more widely available, society is becoming more dependent upon them to perform conventional bookkeeping functions. More important, however, is that as computers become more intelligent they can take on valuable new roles in the service of society. In education, computers constitute a rich new medium for a student's creative expression and experimentation. They can be used to demonstrate laws of physics on a dynamic display screen, to illustrate mathematical principles through the design of algorithms, and to carry on tutorial conversations. In psychology, computer models of mental behavior provide knowledge of how the mind works. In medicine, computers can model physiological and biochemical processes, and both store and deduce large numbers of facts about diseases, drugs, and treatments. In industry, computers can help both in the front office, scheduling activities and monitoring progress, and on the factory floor, directing automatic inspection, materials handling, and assembly systems. Such activities can both increase productivity and improve the quality of the goods produced. In mathematics and science, computers are beginning to function as intelligent assistants to professional scientists, performing such jobs as solving and simplifying symbolic equations, analyzing chemical compounds, and verifying the correctness of simple computer programs. As novel sources of information, amusement, or artistic experiences, the potential for us to benefit from thinking computers is limited only by our imaginations.