Genghis and the Architecture of Decentralised Competence

In 1989 Rodney Brooks and his colleagues at the MIT Artificial Intelligence Laboratory built Genghis, a six legged walking robot approximately thirty five centimetres long. It traversed rough terrain, recovered its footing on uneven ground, and oriented toward a person detected by infrared. Its significance lies in what it did not contain. Genghis maintained no central model of its environment, computed no global plan, and held no representation of its own gait. Its behaviour was produced by a network of fifty seven simple finite state machines distributed across the legs, and the component that coordinated them ran on a single eight bit processor.

Removing the model and the planner

Brooks had spent the preceding years arguing against the prevailing architecture for autonomous robots, in which a system senses the world, constructs an internal model, plans over that model, and then acts. He observed that a single pass through that cycle could require minutes, by which time the modelled situation had changed. His alternative, the subsumption architecture, removed the model and the planner and connected sensing to action through a set of layered behaviours.

The gait is not stored in the system. It is a property of the system in operation.

Each layer implements a simple competence. A lower layer maintains posture; a layer above it advances individual legs; a further layer balances load across the legs in contact with the ground; a higher layer orients the machine toward a target. Higher layers coordinate the lower ones only by suppressing their inputs and inhibiting their outputs. There is no central controller and no shared representation. The walking gait was not programmed. Each leg runs an independent local controller, and the coordinated gait that carries the body over uneven terrain is an emergent consequence of those controllers interacting through the shared mechanical body and the ground. The gait is not stored in the system. It is a property of the system in operation.

The return of the rejected architecture

Stated as a claim about the location of competence, this remains significant. Robust low level behaviour does not require a central controller, and it does not require an internal representation of the task. Brooks summarised the position in the assertion that the world is its own best model, by which he meant that for behaviour grounded in physical dynamics it is more reliable to sense the current state of the world than to maintain and update an internal copy of it. The most widely deployed embodied system in history, the reactive floor cleaning robot, is a direct commercial descendant of this architecture. For accuracy it should be recorded that Brooks rejected central symbolic representation specifically, not representation in all forms; later work in the same tradition reintroduced limited, behaviourally grounded representations.

The contemporary relevance is that the field has, in large part, returned to the architecture Brooks rejected. The dominant pattern for building an autonomous agent places a large language model at the centre and assigns it sensing, modelling, planning, and acting, which is the sense model plan act cycle restored with the model and planner unified in one component. The characteristic failures of these systems, in grounding, in latency, and in the obsolescence of a plan that was correct a moment earlier, are the failures Brooks identified in the original architecture. The prevailing remedy is also instructive: a fast reactive layer is placed beneath a slower deliberative one, with the planner adjusting the parameters of the reactive component. This two layer organisation is not new. Brooks diagrammed it in 1991. The durable result is not that reactive control superseded deliberative reasoning, but that competent autonomy appears to require both, with a reactive layer maintaining contact with physical reality beneath whatever reasoning operates above it.