SHRDLU was a computer program that responded to natural-language commands to move virtual objects, such as blocks, in a virtual world. Today, robotic arms (in the real world) are often programmable and used in assembly lines and bomb disarmament.

1971

SHRDLU

Imagine spending your life in a simple universe consisting of colored objects, like pyramids and cubes, that can be moved around at your will. This is the world of SHRDLU, developed in 1971 by computer scientist Terry Winograd (b. 1946). The SHRDLU program translated natural-language com-mands such as “Will you please stack up both of the red blocks and either a green cube or a pyramid” or “Find a pyramid that is taller than the one you are holding, and put it into the box” into physical actions. One could also ask questions about the history of this universe (e.g., “Did you pick up anything before the cube?”). Internally, it used the LISP programming language and employed simple graphics output to show a simulation of the world that could be manipulated by a virtual robot arm.

In the preface of his 1971 PhD thesis on SHRDLU, Winograd wrote: “Computers are being used today to take over many of our jobs . . . and perform routine office work. . . . But when it comes to telling computers what to do, they are tyrants . . . and act as though they can’t even understand a simple English sentence.” The name SHRDLU comes from “etaoin shrdlu,” the approximate order of frequency of the twelve most commonly used letters in English. To carry out commands, the program had subsystems that parsed language and semantic processing systems that made logical deductions. It had a procedural problem solver that could determine how to carry out the commands, and it kept track of the world, knowing the relative positions of objects.

For its time, SHRDLU was considered a big achievement in natural-language processing. It even had a simple memory: if you told it to move the red ball, and later referred to the ball, it could assume you meant the red one. It also knew what would actually be feasible. For example, it “understood” that the tops of objects must be cleared before new objects could be stacked on top. However, despite its remarkably natural operation, SHRDLU was limited in that it could not learn from errors.