A history of AI & robotics

There is no intelligence here, only choreography. Elektro's "personality" is a script and a record player, which is worth holding onto: a robot performing human-like behaviour is not the same as a robot understanding anything.

Every neural network on this timeline — AlphaGo's, GPT-3's, every vision–language–action model — descends from this simplification. The idea that intelligence might be built from small, uniform, connected units is eighty years old.

The claim and the hedge both still show up in coverage of every model on this page: bold about eventual capability, uncertain about who — or what — is really the audience for it.

No planning, no model of the world — just a feedback loop between two sensors and two motors, and yet it reads as purposeful. It's an early example of a pattern that recurs throughout this page: simple rules can look like intention from the outside.

"75% reliable, and only under specific conditions" could be the honest footnote on several entries much later in this timeline. Stated failure rates are not a modern invention — they're just more often left out now.

"Search over a space of possible moves" is still the core idea behind a chess engine, a Go-playing network, and a language model choosing its next token. The vocabulary is decades old even where the machinery isn't.

LISP became the default language of AI research for decades because that self-inspecting structure suited the field's own ambition: software that reasons about symbols, including its own. Later entries on this page swap symbols for learned weights, but the ambition is the same.

Quicksort itself has little to do with intelligence — it's here because it came out of the same research culture, and its Computer History Museum entry sits in the same timeline. Worth including honestly rather than stretching it into an AI story it isn't.

Assistive robotics — building for a person's needs rather than a factory's throughput — is treated as a fairly recent framing. It's actually as old as computer-controlled arms themselves.

Not every approach on this timeline survives contact with the problem it was built for. Precision and repeatability, not just novelty of design, is what determines which robotics ideas from any era actually stick around.

SHRDLU could genuinely reason within its toy universe in a way ELIZA couldn't fake — and still couldn't leave that universe. The gap between understanding a closed world completely and handling the open world at all reappears constantly further down this page.

This is the point where "robot arm" stops being a one-off lab build and starts being something a factory can order. Research-to-production is a much longer road than research itself, and it runs through here.

Asking a database a question in plain English, rather than a query language, is a problem this field has now been working on for over fifty years — LUNAR just had a much smaller database.

Looking to biology for a mechanical answer, rather than a computational one, is its own long-running thread in robotics — a reminder that not every hard problem here gets solved by adding more intelligence.

Synthetic speech generated from a model, instead of spliced from recordings, is the direct ancestor of every voice assistant on this page. It arrived first inside a children's toy.

A quieter, mechanical kind of progress runs alongside the AI story on this page: better actuators and less friction make every subsequent control algorithm's job easier, whether or not that algorithm involves learning at all.

Industrial robotics was already a global, cross-licensed business by 1982 — component designs, control languages and manufacturing partnerships crossing borders well before "supply chain" was the term anyone used for it.

An autonomous patrol robot for security is a pitch that keeps recurring in robotics, in almost the same words, roughly once a decade. Whether the sensing and judgement behind it are actually good enough is the part that changes.

Building hardware around a specific computational need, rather than adapting general-purpose machines to it, is exactly what today's AI accelerator chips do for neural networks. The strategy is older than the workload it's now built for.

Deep Thought is a direct predecessor of IBM's Deep Blue, eight years later — the same lineage of brute-force search that this timeline's "why it matters" notes keep coming back to when discussing AlphaGo and everything after it.

A remotely piloted aircraft is not an autonomous one, but it normalised the idea of an unmanned platform making armed decisions at a distance — a policy and security question that autonomous robotics still hasn't fully resolved.

"Gradually learns your language" is a scripted timer dressed up as adaptation. Distinguishing a genuinely learned behaviour from a pre-programmed one that only looks like learning is a question this page keeps needing to ask.

ASIMO is the clearest pre-2015 preview of this page's second half: a humanoid body paired with perception and language, decades before Gemini Robotics or GR00T put a learned model behind the same combination.

The Roomba is a deliberately, successfully unintelligent robot: reflexes instead of reasoning, and a commercial hit precisely because that was reliable enough. Not every useful robot needs a model of the world.

Autonomy at Mars-communication distances — minutes of lag each way — forces a rover to make real decisions on its own. It's a much harder version of the same independence every later autonomous system on this page is judged on.

The field's own founders were still alive to watch its fiftieth anniversary — a reminder of how young "artificial intelligence," the named discipline, actually is next to the ideas (McCulloch and Pitts, cybernetics) that came before it.

Watson has to parse an ambiguous natural-language clue and retrieve the right fact faster than a strong human contestant can buzz in — a narrower, more theatrical cousin of the question-answering problem every large language model since has tried to answer more generally.

This is the same year the rest of this timeline's "learning to learn" era begins — DARPA's robots, AlphaGo. The optimism and the warnings about capability start arriving in the same breath, and haven't really separated since.

This is a landmark in learned decision-making. The board, however, has fixed rules and discrete moves. Bringing learning into a moving, uncertain physical world is a different challenge.

Simulation offers a way to gather practice without putting every training attempt through physical hardware. Varying the simulated world helps prepare the controller for imperfect knowledge of the real one.

The useful shift is toward reusing a broadly trained model. Robotics will explore a related idea: adapting a shared model to a new task instead of learning every behaviour from scratch.

Natural language becomes a practical way to interact with a general-purpose model. The launch also documents unreliable answers: fluent conversation does not guarantee an accurate understanding of the world.

A robot can use concepts learned beyond its own demonstrations, including unfamiliar objects and instructions. Generalisation in these evaluations is evidence of progress, not proof of reliability in every setting.

Researchers can inspect, adapt and compare a shared starting point. Access to a model makes experimentation easier; each new robot and operating environment still needs its own evaluation.

The research asks models to handle new instructions, objects and situations across robot bodies. These are reported capabilities in evaluated settings, rather than a claim that general-purpose robotics is solved.

The ambition is to transfer more capability across bodies and tasks. The work is still presented with performance limits and safety evaluations. Broader capability makes the question of dependable behaviour more pressing.