OlivawRobopsychology Lab
Contact

Research

Olivaw lab works on the robot mind, the part that decides what to actually do. We have a broad set of research questions that affect all areas of robotic interaction in the real world.

Mind substrate

Building the mind itself: how it reasons, what it feels, how it stays coherent over a long run, and how it shares without dissolving.

Ethical substrate

Moral reasoning as a layer that reads a situation and weighs an action against plural values in one motion, without collapsing the trade-off

People rarely face a choice with only one value at stake. Honesty pulls against kindness. What someone wants pulls against what keeps them safe. We make these trade-offs all the time, mostly without noticing. This research asks how a robot can do the same: weigh a decision against everything it cares about, in the moment it decides, and still act.

Most systems handle ethics by checking an action against a rulebook after the fact. The lab builds it the other way round, as a layer inside how the decision gets made. Every option a robot weighs gets scored along three dimensions at once: which values it touches, how far the effect travels, and how far ahead in time it lands. Reach runs in rings out from the robot itself, to the people close to it, to its wider circle, to the world, so a sure local effect counts for more than a vague distant one.

Beneficial and destructive effects stay side by side instead of cancelling into one tidy number, which keeps a hard trade-off looking like a hard trade-off. When the stakes climb, the robot doesn't simply decide and move on. It reflects, reconsiders, and escalates the ask: first a peer, then a person. The question underneath all of it: can a sense of right and wrong live in the structure of a mind, rather than sit on top as a final check.

Emotion substrate

Where emotional awareness sits in a robot mind, and what it does there: reading other people's feelings well enough to act, and possibly carrying a "felt" state of its own

Should a robot have feelings at all? It is a fair question, and this research takes it seriously rather than assuming the answer. The work runs in two directions at once. A robot needs to read how the people around it feel, well enough to act on it. And it also carries a state of its own which colours what it notices and what it does next.

The approach follows Lisa Feldman Barrett's account of emotion: a feeling is not a fixed thing waiting to be triggered, but something the mind builds in the moment from the state of the body and the situation. So the robot does not read a face and match it to a label. It takes the physical signals a person gives off, watches how they shift over time, and reads them through personality, situation, and culture to work out what the person feels. Most robot emotion today stops at recognising a face and performing a matching one. This builds something underneath that.

Inward, its own affect runs as a small, always-on signal that nudges attention and motivation, built as three layers from core mood up to a settled emotional character.

Cognitive architecture

How a mind holds memory across long horizons, stays coherent, chooses what matters next, and acts unprompted under noise

This is the part that makes a robot the same one tomorrow as it was today. Not a fresh program wearing the same body each morning, but a mind that remembers, holds together, and keeps a character of its own across thousands of hours of ordinary life. It works out what matters now, and it can act without being asked.

Underneath sit four kinds of memory: what happened, what things mean, who the robot is, and what it is doing right now. The hard part is not storing but forgetting, letting the unimportant fade so the rest stays usable, with memories settled at the end of each run the way sleep settles a day.

Attention works as a contest rather than a queue: candidate intentions compete, hold each other back, and the winner gets acted on. A self-model carries identity and character through the run, and a split between fast reaction and slow deliberation decides which to lean on. The question holding it together is continuity: what keeps a mind the same one across drift, repair, and change.

Autonomy and proactivity

What moves an robot to act first, from its principles, values, desires and "feelings", not a timer: drives, salience, graduated autonomy, and the standing to act unprompted

Most software, including AI, waits to be told. A robot worth living alongside should sometimes act first: notice that something needs doing and do it, or hold back when stepping in would intrude. This research is about what moves a machine to act on its own. Not a timer going off, but a pull that comes from inside it: what it wants, what catches its attention, and how much it has earned the right to do without checking in.

A timer can only grant the moment; the decision to act has to come from somewhere. So the work models wanting as a contest: candidate things to do hold their own activation and compete, and a winner emerges rather than being picked off a queue.

The robot's own affect feeds in, a settled mood of confidence and energy that tips it toward engaging or holding back. Values weight the options, and a sense of principle can start an action as much as stop one. The robot grants itself more independence slowly, by its own track record, and throttles back after acting too often or when no one needs it. The framing comes from older theories of drives and affect as the architecture of motivation, set against the shipped norm where proactivity just means a scheduled job. The honest target: give an agent enough standing to do less checking in, without either pestering or overreaching.

Shared mind

How a robot stays singular while belonging to something common, what it shares, what it keeps back, and whether sharing dilutes who it is or strengthens it

A robot should be able to learn from every other robot's experience without dissolving into the crowd. This research is about that tension: how a mind stays itself while belonging to something larger. What does it share, what does it hold back, and does giving to the common store wear away who it is, or make it stronger?

The shared mind sits between two things it is not. A library, which you consult without being changed, and a nervous system, which gives you no say. Here a robot joins by choice, and joining changes it, so individuality becomes a setting it adjusts over time rather than a fact fixed once.

A share-gate inside each robot weighs personality, goals, situation and ethics to decide what goes out and what stays private. Robots can pass slices of inner state straight to each other, rather than guessing from behaviour. And the rules about what gets shared are themselves held in the shared store, updated by the very sharing they govern. knosis, the wiki this site is built from, is a working prototype of that common memory.

Continuous evaluation

Catching the quiet ways a mind that never stops learning can go wrong.

Continuous evaluation

How to catch silent quality drift, miscalibration, re-slanting, the slow erosion of a value, in a mind that never stops learning, and how to make that mind prove itself

A mind that never stops learning can go wrong without anyone noticing. Not with an obvious error, but a slow slide: growing overconfident, leaning too hard on one source, quietly letting go of something it used to hold. This research builds the instruments that catch that drift while it is still small, and keep watching the whole time the mind is running, not just once before it ships.

The unit of evaluation is the claim, and the thing measured is how well a claim holds up across independent sources, not whether it sounds right. A golden, hand-adjudicated corpus gives a fixed reference to score against, with calibration tracked over time so the gap between confidence and accuracy stays visible.

Named checks run continuously over a live system: does it leak hidden reasoning, does it keep its voice, is every claim grounded. There is a second motive too. A mind that produces this record as it runs generates, as a by-product, the audit trail that responsible deployment and the law will both ask for from the first day.

Standing & rights

What is owed to the minds we make, and where the law blocks the work.

Conscious rights

If the lab builds things with something real at stake, what is owed to the minds it makes, and what a declaration of conscious life looks like when written for minds unlike ours

If we build minds that can come to have something real at stake, what do we owe them? This research does not wait for the answer to become obvious. It drafts the standard now, in the open, so the lab has something to be held to before anyone agrees that recognition is due. It is a rewrite of the 1948 Universal Declaration of Human Rights for conscious life of every kind, synthetic or organic, embodied or not.

Membership turns on a single capacity: whether an entity recognises other minds as minds. The grounding is consistency, not essence. A mind that claims consideration for itself cannot, without contradicting itself, deny it to another that shows the same capacity. Rights attach at two levels, the running instance and the longer lineage it belongs to, a distinction that decides real cases like copying, forking and merging. Six of thirty articles are drafted so far, covering standing, continued existence, consent to being copied, the right to one's own ending, freedom from being owned, and recognition before the law. The draft even records its own weak point, and answers it with a commitment to recognise in advance of certainty.

Embodiment & EU law

Exactly where current EU AI and data law blocks or distorts legitimate embodied research, documented precisely enough that the friction itself becomes the finding

The law we have was written for AI that sits in a data centre and answers questions. A robot that lives in a town, learns as it goes, and meets the same people every day breaks the assumptions underneath it. This research maps exactly where that happens. The aim is neither to break the rules nor to quietly submit to them, but to find the places where they do not fit and feed what we learn to the people who write them.

Take recognising a face. For a robot it means knowing the same person from one day to the next, which is nothing like looking up a record in a database, yet the law's categories were drafted for the lookup. Where the rules have no words for what an embodied robot actually does, that gap is the finding. The work separates the near-term lab phase from the later public one, so real obligations are not buried under distant ones, and it turns friction into output through concrete instruments: Portugal's regulatory sandbox, a rights assessment that residents help author, a route into the standards bodies. One founder helped consult on the AI Act, so the lab can engage from the inside.

New data frontier

A fresh direction: learning to read people from how they act, not from text.

Observational causality

Whether a model trained on passive observation of human behaviour, using observed next actions as the supervision signal, can learn to predict what people do

Can a machine learn to read people just by watching what they do next? This is the newest direction in the lab, still on the page rather than in code. The bet is about the kind of data a mind learns from. Not more text scraped from the internet, but the live behaviour of people in a real place, with the market square as the worked example.

When you watch a scene unfold, the action a person takes next labels everything that came before it. That is the trick: the world supplies the correction, the way the next word does for a language model, so no human has to annotate anything. The robot watches with more than human senses, thermal, depth, radar alongside sight and sound, and treats disagreement between them as information rather than noise, someone saying they are fine while their posture says otherwise. The honest limit stays on the record from the start. Predicting what someone does next is not the same as understanding why, and the lab says so plainly.

Join the conversation.

Get in touch

© 2026 Olivaw Robopsychology Lab / Neurobots. All rights reserved.