Early robots tried to think before moving, and the reaction against that shaped the field

A pipeline that seemed obviously correct

The mobile robots built at research institutes from the late 1960s onward were organised around a sequence: sense the surroundings, build an internal representation of them, reason about that representation to produce a plan, then execute the plan. Each stage was a separate research problem and each was worked on separately.

The arrangement follows naturally from a view of intelligence as reasoning over symbols. If thinking is manipulating a model of the world, then a machine needs a model, and building one from sensors is simply the input problem. The robots were as much a demonstration of that philosophy as they were engineering.

They worked, in the sense that a robot could be given a goal in a room prepared for it and would eventually accomplish the goal. The eventually is doing a lot of work. Deliberation took a long time, and the world was required to hold still while it happened.

The bottleneck turned out to be perception, not reasoning

The planning was the part everybody expected to be hard, and the planners produced were genuinely capable. What proved intractable was building the representation in the first place. Turning camera images into a reliable description of where objects are and what they are was far beyond what could be done, and remained so for decades.

Because the pipeline was sequential, a weakness at the first stage propagated. A plan reasoned impeccably over a mistaken model of the room produces confident, precise, wrong behaviour. The system had no means of noticing, since its only account of the world was the one that was wrong.

This is the general observation later summarised as the point that tasks people find difficult are computationally easy and tasks people find trivial are computationally very hard. Chess is easier than walking across a cluttered room, and the reasons are structural rather than a matter of effort.

The reaction proposed doing without the model

By the mid-1980s a body of work argued that the internal model was the problem rather than the solution. Its central claim was that the world is its own best model and should be consulted directly rather than represented, and that useful behaviour could be built from layers of simple sense-act couplings with no central reasoning at all.

Robots built this way were startling. They moved continuously, reacted immediately, coped with things being moved around them, and did so with computation that was trivial by comparison. Behaviour that looked purposeful emerged from the interaction of simple rules with a complicated environment.

The provocation in the argument was philosophical as much as engineering. It held that representation and reasoning were not merely inefficient but the wrong account of how intelligent behaviour is produced, which put it in direct opposition to the field’s founding assumptions.

Neither position won and both were partly right

The reactive approach produced excellent behaviour and struggled with anything requiring a goal held over time, or a plan that could not be discovered by acting. Deliberation exists for a reason: some problems must be solved before acting because acting to find out is too expensive or irreversible.

What emerged in practice were layered architectures with fast reactive control underneath and slower deliberation above, each operating at the timescale it suits. That is now unremarkable engineering, and it was arrived at by a decade of argument that both sides describe as having been won by their own position.

Commercially successful robots have generally been those that reshaped the environment rather than the reasoning. Guaranteeing what a robot will encounter removes most of the perception problem, which is why industrial arms in fixed cells worked long before anything worked in a house.

The same argument is being had again

Contemporary systems trained end to end from data to action, with no explicit world model, are in the same relation to model-based approaches as the reactive robots were. The arguments made for and against them rhyme closely with the earlier ones, generally without acknowledgement.

The unresolved question is whether a system needs an explicit representation of the world to plan reliably within it, or whether a sufficiently trained mapping from perception to action does the job implicitly. Serious researchers hold both positions, and evidence is offered for both.

What the earlier round suggests is that the answer will be a division of labour rather than a victory, and that whichever approach appears to be winning will be found to have absorbed the useful parts of the other. That is a pattern, not a prediction, and it may not repeat.