About Olivaw

Over the next decade, robots will move into the parts of life that people actually inhabit: the streets, the shops, the home, alongside the work they already do in factories and fields. The hardware for it has nearly caught up. Robots can move and balance and grip, they can see and recognise what they are looking at, and they can plan a route and follow it. They still lack a mind. By a mind we mean the thing that holds together from one moment to the next, that remembers, that carries a character, and that works out what to do when the situation gives no clear signal, when the rule does not quite fit, or when someone needs something nobody thought to specify in advance.

Olivaw works on that part.

What the lab works on

Asimov gave the word robopsychology its first outing in 1950. We use it more or less as he meant it: the discipline of building a synthetic mind and staying responsible for how it grows up.

A robot has several layers. At the bottom sit the reflexes that keep it upright and let it handle objects. Above those, the perception that turns sensor readings into a picture of the world. Above that, the planning that turns a goal into a sequence of actions. Each of these layers occupies serious groups doing serious work, and a good deal of money and talent now flows into them. Olivaw works on the layer above all of them, the cognitive one: memory, motivation, a sense of self, the reading of other minds, the handling of a social situation, and ethics. This layer decides what the machine actually does, and it has had rather less attention than the ones beneath it.

The mind the lab builds should run on more than one kind of body, so the work stays with the cognitive layer rather than binding itself to any single manufacturer’s hardware.

Why it has to grow in a real place

Robotics has a familiar and dependable way of working: assemble the system, test it under controlled conditions, refine it, release it. It suits the making of a capable machine and it has carried the field a long way. A mind seems to need something that method cannot give it.

A mind grows by meeting consequences, and a consequence has to fall on someone real, in a real place, to count for much. In a simulator a mistake comes back as a number. In a town it comes back as a shopkeeper’s irritation, a small public awkwardness, a person who quietly decides to give the robot a little more room next time. It arrives all at once and tangled together, the way things do in ordinary life, and a mind forms by growing against that kind of tangle. No one can manufacture it. So the lab places its robots in an ordinary working town and treats their life there as the research itself, rather than running a build phase first and a deployment afterwards.

A consequence has to fall on someone real, in a real place, to count for much.

Why now

Two clocks run at different speeds here. Hardware improves from one month to the next, and the foundation models underneath it improve just as quickly, so what robots can physically do keeps extending. Nobody can hurry a mind to match. It forms out of a long accumulation of real experience, and that takes the time it takes; neither money nor extra hands will compress it. Given that mismatch, the lab starts the slow work early, so that some understanding of how these minds should grow has a chance of existing by the time the robots arrive in numbers. The questions will find answers one way or another, whether deliberately and in the open or hurriedly and after the fact. The lab would rather work them through with care, and with the people who will live alongside the results.

Two clocks run at different speeds here.

Two labs

The programme runs in two places. The mind is grown in Caldas da Rainha, a town of around fifty thousand people in the Oeste region of Portugal, and it is turned into something other people can use in Abu Dhabi. The two are not a headquarters and a branch; each does something the other cannot.

Caldas is where the consequences are real. The size of the town matters more than it first appears. A village would give a robot too little to grow against, the same few faces and situations coming round again and again. A large city would give too much and notice too little; the robots would disappear into it, and the place would never really register them or answer back. A town holds enough ordinary complexity to learn from, a daily market, a hospital, schools, commerce and agriculture, regulars and strangers, while staying small enough to feel the robots arrive. Over recent years many people have moved there from elsewhere, so that about a fifth of residents now come from outside Portugal, across a range of languages and backgrounds; a mind learning to read people has more to learn from in that mix than in a uniform population. The town also has a long habit of absorbing the unfamiliar, having grown up around a thermal hospital that Queen Leonor founded in 1485 and a ceramics tradition that came after. A simpler reason matters too: the founders live there, which keeps the lab answerable to its neighbours.

The town takes part as a partner rather than serving as a backdrop. The work proceeds with the municipality and with the local design school and polytechnic, and the lab builds the governance in from the start: a citizen oversight board with real authority over where and when the robots operate, straightforward ways for residents to opt out, and clear safety arrangements, including the means to stop a robot, that stay outside negotiation.

Abu Dhabi is where the work becomes usable by someone else. A mind that has been grown does not arrive as a file inside a robot: part of it runs on the machine and part of it runs remotely, and that infrastructure lives there. So does the repeated testing that shows a behaviour holds a second and a hundredth time rather than once, which a town cannot produce and a built environment can. And so does the floor where the people who will put the layer into their own hardware come to see it work and talk to the people who keep it running. A research result nobody else can operate is not much of a result.

What the lab builds on

The lab does not begin from nothing. Work on two of the harder parts of the problem has already come a long way.

The first reads emotion. It takes the physical signals a person gives off, follows how they shift over time, and interprets them through cultural and situational context rather than reading a fixed expression off a face. It rests on Lisa Feldman Barrett’s account of emotion as something the brain constructs, rather than on the older idea of a small set of universal, hard-wired emotions. Several years of development stand behind it.

The second lets a robot weigh a decision against its values in the moment of deciding, rather than checking the action against a rule once it has acted. It treats values as things that carry weight, that reach different distances from the act, and that fade over different spans of time, and it keeps the competing considerations visible instead of collapsing them into one score, sending the hardest cases to a person. It too has taken several years.

Behind both lies a longer line of work on agentic architectures for disembodied AI: software that plans, holds a goal across many steps, keeps a memory, and decides what matters next. Much of what that established about memory and about staying coherent over a long run carries over into the embodied setting.

None of this stands apart from the wider field. The lab draws on decades of research into cognitive architectures, machine memory, and machine ethics, and it works alongside the groups building robot bodies and the labs building spatial world-models. Its own contribution stays deliberately narrow: the cognitive layer, developed by letting people and AI grow together in a real place. That marks a difference of method, not a claim to have reached somewhere the rest of the field has not.

An open question

One question sits underneath all of this, and the lab would sooner raise it than wave it away. How could anyone tell whether a robot has come to have a self, as opposed to a very convincing imitation of one? From the outside, good pattern-matching and selfhood look much the same, and you could point to no component inside the machine and call it the self.

The lab’s working answer rests on time. Give a pattern-matcher and a genuine self the same year of real consequence, and they start to come apart on their own. One accumulates nothing in particular. The other ends up with a history that belongs to it alone, and with people it could let down. A self, on this reading, amounts to something that has lived long enough to have something to lose. And if the lab does end up making things that can have something to lose, it will owe them some consideration. Quite what the lab would owe them, no one yet knows. The lab means to keep the question open, and in plain view, as the work goes on.

A self amounts to something that has lived long enough to have something to lose.