Summary
Brooks’ paper is a direct attack on the classical AI program: the assumption that intelligence requires building an explicit internal model of the world, manipulating symbolic representations, and planning before acting. Brooks argues this is wrong both as engineering and as biology, and demonstrates that complex, adaptive behavior can emerge from reactive systems with no internal representation at all.
The argument against representation
Classical AI (McCarthy & Hayes (mccarthy-1969-some-philosophical-problems-from-the), Minsky’s frames (minsky-1974-a-framework-for-representing-knowledge)) assumed:
- Perception → build world model
- Reason over world model → generate plan
- Execute plan → act
Brooks identifies three failure modes:
- The representation bottleneck: building a sufficiently complete world model is computationally intractable
- The frame problem: updating a representation when the world changes requires specifying what didn’t change — an open-ended problem
- The execution gap: plans assume the world holds still; real robots operate in dynamic, unpredictable environments
His alternative: “the world is its own best model.” Instead of representing the world, use the world directly. Sensors read current state; actuators affect it. No intermediate representation.
Subsumption architecture
Brooks’ engineering solution: layer behaviors hierarchically by priority.
- Layer 0 (lowest): avoid obstacles — always active, highest priority
- Layer 1: random wandering
- Layer 2: explore toward distant objects
- Layer 3: identify and approach objects of interest
- Higher layers: pick up, return to home, etc.
Higher layers subsume (override) lower layers only when their conditions are met. Each layer is a finite state machine that reads sensors and drives actuators. No layer has access to the others’ internal state; communication is only via the shared sensor/actuator space.
Key result: Brooks’ robots (Allen, Herbert) exhibited sophisticated-looking behavior — navigating cluttered offices, collecting cans — with no world model. Observers consistently attributed intentionality to behavior that had none.
Embodiment and situatedness
Brooks coined the terms with Pattie Maes: embodied AI (the body’s physical interaction with the world shapes intelligence) and situated AI (intelligence is always relative to a context, not abstract). These became the foundation of behavior-based robotics and later influenced cognitive science (Lakoff’s embodied cognition, Varela’s enaction).
The contrast with LLMs is sharp: LLMs are disembodied — they process text tokens with no sensorimotor grounding. Brooks would argue this limits them to symbol manipulation, not genuine understanding.
Connection to ALife
If von Neumann asked “what is the minimum complexity for self-reproduction,” Brooks asked “what is the minimum complexity for intelligent behavior?” His answer — reactive, embodied, no representation — influenced Langton’s ALife program: emergent complexity doesn’t require top-down design, only local rules and environmental feedback.