Textbook of AI

The blocks world is a simulated environment of coloured blocks and pyramids on a virtual flat table, used as a testbed by an entire generation of MIT and Stanford AI researchers in the 1960s and 1970s. With a small number of objects, a small number of relations (on, under, in front of, to the right of, …) and a small number of actions (move, stack, take, place), the blocks world supported a substantial amount of foundational AI work.

What was built in it

Successive blocks-world systems showcased advances across nearly every AI subfield:

• Natural-language understanding, Terry Winograd's SHRDLU (1972) understood natural-language commands such as "find a block which is taller than the one you are holding and put it into the box," held a dialogue about the scene, and answered questions about its own actions.
• Planning, Gerald Sussman's HACKER (1973) learned plans from failures; the STRIPS planner (Fikes and Nilsson, 1971) was demonstrated on blocks-world tasks; the Sussman anomaly is a canonical blocks-world example of plan-interaction failure.
• Computer vision, Larry Roberts's 1963 thesis extracted polyhedral structure from images of blocks; David Waltz's line-labelling (1972) constraint-propagated line junctions to recover scene geometry; David Marr's early work used blocks-world scenes.
• Machine learning, Patrick Winston's 1970 thesis learned structural concepts ("arch," "tower") from positive and negative examples expressed as blocks-world descriptions, including the celebrated near-miss learning strategy.
• Theorem proving and knowledge representation, frame-based descriptions of blocks-world scenes were a standard testbed.

The intellectual bet

The shared assumption was that competence in the blocks world would scale to competence in the real world. If a system could understand "the small red block on top of the green pyramid," it would surely generalise to "the small red car parked behind the lorry." The blocks world stripped away perceptual and physical messiness so researchers could concentrate on representation, reasoning and language.

The bet did not pay off. Every successful blocks-world system relied on a great deal of hand-crafted, world-specific knowledge, closed-world assumptions, perfect perception, deterministic actions, complete and consistent ontologies. None of these survived contact with the real world, and the limits of the approach became one of the recognised lessons of the 1970s and the first AI winter. Hubert Dreyfus's What Computers Can't Do (1972) had warned that micro-worlds would not scale; the failure to extend SHRDLU beyond its 1972 demo lent his critique force.

Legacy

The blocks world remains a useful pedagogical environment for teaching planning, search, knowledge representation and rule-based systems. It survives in AI textbooks (Russell & Norvig, Nilsson) and undergraduate coursework. Modern echoes appear in:

• Embodied LLM benchmarks, text-based blocks-world manipulation tasks now serve as evaluation suites for tool-using language models.
• Robotic manipulation, pick-and-place benchmarks (Ravens, RLBench) often use simplified block environments.
• Reinforcement-learning task suites, many gridworld and stacking benchmarks are direct descendants.

The blocks world is thus both a humbling cautionary tale about the difficulty of generalisation and a continuing source of pedagogical and benchmarking value.