Brooks’ paper is the most direct challenge to classical AI ever written by a working roboticist. Its thesis: the fundamental assumption of AI since McCarthy — that intelligence requires explicit symbolic representation — is simply wrong, empirically falsified by the behavior of insects and practically refuted by the engineering failures of classical planning systems.
The classical AI assumption and why it fails
The GOFAI (Good Old-Fashioned AI) pipeline:
World → Perception → World Model → Planning → Execution → Action
Every step is a bottleneck:
- Perception → World Model: real environments are too complex to model completely. Any omission is a potential failure mode.
- World Model → Planning: planning over complete world models is computationally intractable (PSPACE-hard for many problems).
- Planning → Execution: the world changes while the plan executes. Real robots fail constantly because their plans assume a static environment.
The frame problem (McCarthy & Hayes (mccarthy-1969-some-philosophical-problems-from-the)): when something happens, how do you update your world model? You must specify what didn’t change — an open-ended problem since almost everything is unchanged by any given action. Classical logic has no satisfactory solution.
Subsumption architecture
Brooks’ solution: eliminate the world model entirely. Replace it with a layered set of behaviors — each a finite state machine that reads sensors directly and writes to actuators:
Layer 0: Avoid obstacles ←── always active
Layer 1: Wander randomly
Layer 2: Explore toward openness
Layer 3: Approach interesting objects
Layer 4: Acquire objects
Higher layers subsume lower ones: when layer 3 is active, it overrides layer 1’s random wandering. When layer 3 deactivates, control falls back to layer 1. No global state, no world model, no planning. Each layer is independent and testable.
“The world is its own best model”: Why build an internal representation of the environment when the environment is always right there, readable by sensors? Any internal model is a simplification and therefore wrong; sensors give the ground truth.
Experimental evidence
Brooks’ robots (Allen: box-avoiding mobile robot, Herbert: soda can collector) navigated real cluttered environments, interacted with humans, and exhibited apparently goal-directed behavior — entirely through reactive layers with no planning. Human observers consistently attributed intentionality and intelligence to behavior that was purely reactive.
This was a direct empirical refutation of the implicit assumption that sophisticated behavior requires sophisticated planning.
Embodiment and its implications for LLMs
Brooks coined embodied AI: intelligence is not in the head alone but in the coupling of body, brain, and environment. A disembodied intelligence — one that reasons about the world without being in it — will have fundamental limitations that an embodied one avoids. The body’s interaction with the world offloads cognitive work: you don’t need to remember where the cup is if you can just look.
This is a sharp critique of LLMs, which are maximally disembodied — they process token sequences with no sensorimotor loop, no proprioception, no environmental feedback. A Brooks-ian would argue this places a fundamental ceiling on the kind of intelligence they can achieve, regardless of scale.
Connection to ALife and RL
Brooks’ reactive agents are the physical instantiation of ALife’s core claim: complex behavior emerges from simple local rules without central control. The subsumption architecture echoes Langton’s edge-of-chaos argument (langton-1990-computation-at-the-edge-of): ordered behavior emerges not from a plan but from dynamics.
Reinforcement learning (williams-1992-simple-statistical-gradient-following-algorithms-for) is in some ways a synthesis: RL agents operate reactively (a policy maps states directly to actions, like Brooks’ behaviors) but learn that policy from interaction (addressing Brooks’ limitation: reactive systems can’t improve beyond their programmed behaviors). Modern deep RL + robotics is the attempt to combine Brooks’ embodiment insight with Holland’s learning-from-selection insight.