Blocks (cubes)
Rectangular solids that can be stacked on one another, placed in the box, or set on the table. The workhorses of the scene, they support other objects and can be supported in turn.
The blocks world
SHRDLU could not see the real physical world. It looked onto a tabletop, and that tabletop existed only inside the machine. The famous dialogue takes place in a simulated scene of coloured blocks and pyramids, a box that can hold things, and a table that holds everything else, manipulated by a simulated hand that picks up, moves, stacks and lets go. On the DEC-340 display the scene appeared as a clean line drawing, an axonometric sketch of geometric solids that a viewer could read at a glance. Later work at the University of Utah added three-dimensional colour rendering, but the world being rendered was always the same small, enclosed arrangement of objects on a surface.
It is tempting to treat this smallness as a limitation, the regrettable scaffolding of an early system that would later be removed. The argument of this page is the opposite. The smallness is crucial for SHRDLU's being able to work at all. Because the world contained only a handful of object types and a handful of relations, every fact about it could be hand-built in advance, and the program could be given the means to resolve every question the world could pose. What looks like comprehension is, read closely, the achievement of a domain engineered to contain nothing the program had not already been taught to handle.
WHAT WAS IN THE WORLD
What was in the world
The objects manipulated in the canonical dialogue are blocks (cubes), pyramids, a single box that functions as a container, and the table that serves as the ground surface. The blocks and pyramids come in three colours, red, green and blue, and in contrasting sizes, big and small, tall and short. From this handful of types the entire demonstration is organised to show how a limited domain can be programmed as explicit knowledge representation into computer source code.
Pyramids
Solids that can be picked up and placed but not stacked upon, since nothing sits stably on a point. The program knows this constraint and reasons with it when asked to build.
The box
A container rather than a buildable solid. Objects can be in the box as well as on surfaces, which gives the world its one genuine spatial relation of containment.
The table
The ground of the scene, the surface on which everything ultimately rests. It anchors the recursive chain of support that the program traces when it explains why it did what it did.
The relations between objects in this block world were also limited. The program could reason about an object being on another, in the box, that one block supports another, that an object is behind another in the scene, and that one solid is bigger, narrower or wider than another. These predicates exhaust the spatial logic of the world. There is no notion the program lacks the means to compute, because there is no notion in the world that the predicates do not cover.
Manipulation is carried out by a simulated hand, an arm and gripper that can grasp a single object, move it through the scene, place it on a surface or in the box, and release it. The arm and the blocks alike are data structures the program updates as it reasons. SHRDLU's object lexicon admitted further shapes, including cones, and some accounts add balls, but the canonical demonstration exercises only blocks, pyramids, the box and the table, in red, green and blue.
THE MICROWORLD IDEA
The microworld idea
The blocks world is the paradigm case of what the MIT Artificial Intelligence Laboratory called a microworld, the research programme associated with Marvin Minsky and Seymour Papert. A microworld is a domain deliberately made small enough that all the relevant knowledge can be hand-built into the program. Rather than confront the open, untidy world in its full ambiguity, the researcher carves out a closed region of it, axiomatises that region completely, and builds a system that operates with apparent fluency inside the boundary.
SHRDLU's world is describable in roughly fifty words, and its working vocabulary is of a comparable order. This is a design feature of the block world and makes it possible to reduce the breadth of knowledge that needs to be programmed considerably. Every object type, every colour, every spatial relation and every legal action was specified in advance and represented explicitly in the program's procedures. Nothing in the world arrives unannounced, and nothing the user can refer to lies outside what the system has already been told.
The wager of the microworld programme was that you could master a small closed world completely and then scale outward, that intelligence built for the tabletop would generalise to the room, and the room to the world.The microworld wager
The expectation was that understanding achieved under controlled conditions would be the seed of understanding in general, and that the techniques honed on blocks would extend, with effort, to richer and less tractable domains.
WHAT THE WORLD LEAVES OUT
What the world leaves out
Read as a cultural text, the blocks world is most revealing in what it excludes. It is closed, frictionless and fully known. There is no weather, no wear, no object that behaves unexpectedly, no perception that can be mistaken. The hand never fumbles, the blocks never chip, and the table never runs out of room in a way the program has not anticipated. The apparent intelligence of SHRDLU is, on close reading, the intelligence of a world built to contain no problem it cannot already solve.
Crucially, there is no real perception here. The arm and the blocks are simulated, not the output of robotic vision, so the hard problems of seeing, of segmenting a noisy scene into objects, of recognising a block as a block under shadow and occlusion, simply do not arise. The world is delivered to the program pre-interpreted, already designed and written into the very categories the program reasons with. What we are invited to read as comprehension is purchased by this radical restriction. The world has no ambiguity that the program has not been given, in advance, the means to resolve.
PersonWHAT DOES THE BOX CONTAIN?
SHRDLUTHE BLUE PYRAMID AND THE BLUE BLOCK.
the question can only ever name objects the world already holds
The question can be answered exactly and confidently because the box can contain only objects already in the world, described only by attributes the program already tracks. There is no possibility of a surprising answer, a partly seen object, or a referent the system cannot pin down. This is the problem identified as a scaling problem. A method that succeeds by knowing everything in advance has no obvious route to a world where everything cannot be known in advance, and the leap from the tabletop to the real-world is not something that was obvious, or, as it turned out, possible, through symbolic forms of expert system.
The canonical dialogue shows the world being queried in natural language with fluency, and the program shows how that fluency was assembled from a parser, a procedural grammar, semantic analysis and a deductive problem-solver.