Source
Brooks (1991), “Intelligence without representation”, LIT-436 (Artificial Intelligence 47:139–159, written in 1987), read in full in NOTE-351 from the copy Brooks posts.
What was actually shown
The architecture (§6.2). Each layer is a fixed network of finite-state machines with “a handful of states, one or two internal registers, one or two internal timers”, passing 1-bit or 24-bit messages. A higher layer subsumes a lower one by suppression (a new wire replaces an input for a set time) or inhibition (a new wire blocks an output for a set time). There are no variables, no rule matching and no shared data.
The robot. On Allen, three layers: avoid (twelve sonars, collide, feelforce, runaway, turn, forward); wander (a random heading every ten seconds or so, summed with repulsion); explore (look, stereo, path-plan, steering around the obstacles the lower layer avoids). The robots ran “for well over a year”. Six layers ran in simulation.
What follows by construction (§5). “Not by design, but rather by observation”, there is no central representation. Each layer has “its own implicit purpose (or goal if you insist)”, checked continuously against the world, and “there need be no explicit representation of goals that some central (or distributed) process selects from”. “Just as there is no central representation there is not even a central system” (§5.1). Low layers react fast because they keep no representations, and a change in the world disables some layers, not all.
What could have come out otherwise: the layered robot could have failed to behave purposively in a real, unsimplified environment, or needed an arbiter. As described, it did neither, at three layers.
What this does not say
- That intelligence in general needs no representation. That is the paper’s labelled hypothesis (H), supported by the robots and an evolutionary timeline. The timeline does not show that reasoning is “pretty simple” once mobility is solved.
- That there are no representations in the robot. Brooks declines the word because the system has no variables, rule matching or choice. He concedes that “an extremist might say that we really do have representations, but that they are just implicit”. The claim is about explicit, central, symbolic representation.
- That the robot’s unity is only in the observer. Brooks says coherence emerges “in the eye of an observer”. That is an interpretive assertion, not a result. Whether the whole is a real agent is not argued against.
- That it scales. How many layers before their interactions become too complex, how complex the behaviours can be, and whether learning is possible are posed in §8 and not answered. The fourteen-layer robot is a plan, and learning works only “as an isolated subsystem”.
- That performance was measured. It is described, never measured or compared. “The most reactive real-time mobile robots in existence” is the author’s report.
- Anything about minds or experience. “I have no particular interest in demonstrating how human beings work”, and “no biological significance” is claimed.
Connections
- Minsky’s society of mind (LIT-429) gives, in Brooks’s words, “a similar account of how human behavior is generated”. Minsky’s agents carry memory and representational machinery; Brooks’s layers pass uninterpreted numbers. This is the more austere decomposition, built.
- THEORY-043 treats a mind of mindless agents as one direction of a single functionalist move. This is the mindless direction realised in hardware.
- Goal-directedness without self-representation: Schwitzgebel (LIT-159) counts goal-directed responsiveness, self-monitoring and self-representation among the marks of a possible group mind. These robots have the first and none of the others, so the marks come apart (NOTE-351).
- CLAIM-045 assigns Brooks the sub-question of how coherent behaviour arises without a central executive. This is the result it relies on.
- THEORY-056 holds that emotion regulation is one motive state checking another, with no regulating system above them, and cites this architecture as that structure built.