SHRDLU is an early natural-language understanding computer program developed by Terry Winograd at MIT in 1968–1970. It allowed users to converse with a computer in English, manipulating objects in a simplified "blocks world" - a virtual box containing blocks, cones, and balls. The program combined language parsing, memory, and basic physics to simulate understanding, and was written in Micro Planner and Lisp on a DEC PDP-6 computer with a DEC graphics terminal. Later additions at the University of Utah added full 3D rendering of the world. The name derives from ETAOIN SHRDLU, the letter arrangement on Linotype machines, ordered by frequency of use in English.
Functionality
SHRDLU's core was a language parser that handled user instructions and queries about the blocks world. The world was deliberately simple, with a vocabulary of about 50 words: nouns like "block" and "cone", verbs like "place on" and "move to", and adjectives like "big" and "blue". This simplicity allowed the program to parse sentences and execute actions reliably.
A key feature was its memory, which provided context. Users could refer to previously mentioned objects without repeating full descriptions. For example, after "put the green cone on the red block", a subsequent "take the cone off" would correctly identify the green cone. SHRDLU could also answer questions about its history, such as "did you pick up anything before the cone?"
The program included basic rules about the world's physics, enabling it to reason about possibilities. It could deduce that blocks could be stacked but that pyramids could not, based on its experiences. This allowed it to answer questions like "Can a pyramid support a pyramid?" with "I DON'T KNOW" or "Stack up two pyramids" with "I CAN'T".
SHRDLU also allowed users to define new terms. For instance, one could say "a steeple is a stack which contains two green cubes and a pyramid", and SHRDLU would then understand and answer questions about steeples, even building new ones on request.
A typical session (as documented in Winograd's work) shows the program handling complex instructions, resolving pronoun references, and explaining its actions. For example, when asked "Why did you clear off that cube?", it replied "TO PUT IT ON A LARGE GREEN CUBE", demonstrating a chain of reasoning.
Consequences
SHRDLU was considered a landmark success in artificial intelligence (AI) research, showcasing the potential of natural-language understanding. Its performance led to optimism among AI researchers, who believed that scaling up such systems could achieve general intelligence. However, this optimism faded when later systems faced real-world ambiguity and complexity, which SHRDLU's controlled environment avoided.
The program influenced subsequent work in Artificial intelligence and Natural language processing, though it did not directly lead to modern approaches like Machine learning or Large language models. Winograd's work at MIT CSAIL (then the MIT AI Lab) contributed to the field's foundational concepts, including the importance of context and world knowledge in language understanding.
Technical Details
SHRDLU ran on a DEC PDP-6 computer, a mainframe of the era, and used a DEC graphics terminal for display. The software was written in Micro Planner, a planning language, and Lisp, a programming language popular in AI research. The program's architecture integrated a syntactic parser, semantic interpretation, and a problem-solving component that manipulated the blocks world.
The blocks world itself was a simulated environment with objects like blocks, pyramids, and balls, placed on a table or in a box. The program tracked object positions, sizes, colors, and relationships, enabling it to answer spatial queries and execute manipulation commands.
Legacy
SHRDLU remains a classic example of early AI, often cited in textbooks and historical accounts. Its demonstration of context-sensitive dialogue and planning was ahead of its time, though its limitations were also instructive. The program's reliance on hand-crafted rules and a restricted domain contrasted sharply with later data-driven approaches like Deep learning and Neural networks.
Winograd later moved to Stanford, where he continued research on human-computer interaction. The name SHRDLU itself became a curiosity, derived from the Linotype machine's key order, and is sometimes used as a term in computing.
Despite its age, SHRDLU's principles - such as using memory for context and reasoning about physical constraints - remain relevant to modern AI systems, though they are now implemented differently. The program's success and subsequent decline in optimism highlight the challenges of achieving robust language understanding, a goal still pursued by researchers today.