Rodney A. Brooks (1991), Artificial Intelligence 47(1–3), 139–159 (received September 1987; issue dated January 1991) — DOI-10.1016/0004-3702(91)90053-M. The author’s copy is at https://people.csail.mit.edu/brooks/papers/representation.pdf
Key takeaways
- Build complete Creatures, incrementally, in the real world. “At each step we should build complete intelligent systems that we let loose in the real world with real sensing and real action.”
- A conclusion and a hypothesis. (C): “explicit representations and models of the world simply get in the way. It turns out to be better to use the world as its own model.” (H): “Representation is the wrong unit of abstraction in building the bulkiest parts of intelligent systems.”
- Decompose by activity, not by function. Each layer “individually connects sensing to action”. A new layer is added in parallel to a debugged one, which “continues to run … unaware of the existence of the second level”.
- No centre. “It is only the observer of the Creature who imputes a central representation or central control. The Creature itself has none; it is a collection of competing behaviors. Out of the local chaos of their interactions there emerges, in the eye of an observer, a coherent pattern of behavior.” Brooks adds: “Minsky [10] gives a similar account of how human behavior is generated.”
- The subsumption architecture. Each layer is a fixed-topology network of finite-state machines exchanging short messages over wires, with “no possibility of access to global data … no possibility of global control”. Layers combine by suppression (on a machine’s input) and inhibition (on its output).
- Evidence is modest and stated as such. At most three layers ran on a physical robot, and six in simulation. A fourteen-layer soda-can collector is described as in progress. “Only performance truly counts.”
Standing in the record
Filed on 2026-10-02 for the owner’s bridge between Minsky’s society of mind and Schwitzgebel’s conscious United States. NOTE-351 is the close reading, and it placed the work: Active, on the decomposition side.
It is the nearest thing to a built Society of Mind the bridge has. Brooks cites Minsky’s Society of Mind (LIT-429) as “a similar account of how human behavior is generated”, and his robots are the engineering case. Purposive behaviour comes from many simple parts that do not share representations, and no part is in charge. Singh’s exposition (LIT-444) reports that the Society of Mind itself has never been implemented. Brooks’s architecture is a different and much smaller design, but it is built. Its claim that the whole’s unity is imputed by an observer is the decomposition-side mirror of Schwitzgebel’s telescopic view in LIT-159 (NOTE-131). Brooks reads the unity as the observer’s. The telescopic view reads it as real.
On NOTE-131’s numbering it is neutral on C3, C4 and C5. It says nothing about consciousness, and Brooks has “no particular interest in the philosophical implications of Creatures”. His parts are deliberately mindless and his wholes are insect-level at best. It bears on the seam in two narrower ways, both set out in the note. First, its Creatures have goal-directed behaviour with “no explicit representation of goals” and no self-representation. That separates two of the marks Schwitzgebel lists for the United States. Second, it is an architecture whose layers coordinate mostly through the world, not through each other. Under IIT 3.0 (LIT-445) that kind of coupling does not count as integration.
It is not the “Brooks (1986)” that NOTE-131 lists in Schwitzgebel’s functionalist tradition beside Block and Lycan. That is a philosopher’s group-mind case. Crossref holds a D.H.M. Brooks, “Group minds”, Australasian Journal of Philosophy 64 (1986), 456–470, but whether that is Schwitzgebel’s reference was not checked here.
Boundary. The paper carries an instruction for AI practice, in its
methodological maxims (§6.1). Test in the real world, never in a
simplified one. Build and debug one layer at a time, so the only thing
left to vary is the new layer. Every step must be a complete system. That
is robotics and AI methodology, not machine-learning practice as such. But
the anthology’s agents-and-environments topic, “percept–action loops …
and what acting rather than only predicting changes”, could hold it, so it
is flagged anthology-candidate under ADR-013. The anthology does not hold
it now. If it is filed there, the anthology would read it for agent design
and this record for the part–whole question.