When AI leaves the screen.

AI learned to talk; now it is learning to move

In 2023, researchers at Google DeepMind described RT-2, a robot model trained on robot data and on text and images from the web. Because it had learned from the web, it could turn instructions it had not met in its robot training into actions. In 2025 the same lab introduced Gemini Robotics, a model that controls robots directly, copes with objects and places it has not seen, and can pick up a new short task from about a hundred demonstrations.

Language models taught AI to talk. Models like these are teaching it to act in a room. That, we think, is the part of AI that will reach the home, and it is the part we are building for.

What a home needs from a robot

Not a butler. A home needs something small and gentle that knows the place: the hallway, the stairs, the cat’s favourite chair. It should come when you ask, go and look when you cannot, and keep out of the way the rest of the time. Our idea of a companion is one you know: it knows your home and your routines, and you know exactly what it will and will not do. Without the second half, the first is unsettling.

That is also why most of the talking happens on your hand or at the dock, where it is quiet. Flying is for going, looking and finding.

Can a machine be company?

Researchers have asked this for years, mostly with older adults. A 2021 review of 29 studies found that in most of them, robots or computer agents improved at least one measure of loneliness. They helped in two main ways: as company in their own right, and by prompting contact with other people. There was a cost too: some people were sad when the robot was taken away.

Studies of chatbots point the same way, and add a warning: relief in the moment, and worse outcomes for the heaviest users over weeks. Our reading is that a companion should bring people closer to each other, not stand in for them: a reminder to call your sister, a message from someone you chose, a quiet evening with the phone in another room.

The rules come first

• The AI proposes; tested code decides. Whatever model you connect can suggest a flight; the safety guard, which does not learn, decides what the motors do: the AI proposes, the code decides.
• Distance. Alkyon is designed to keep at least 1 m from people and pets, and to slow down when someone is in the room.
• The camera stays home. Its images are processed on the dock, in your home.
• Your model, your choice. You can connect the AI model you like; it gets what a request needs, not the run of the house.
• One press stops it. Pressing the ring twice stops the robot.

Where we are

Lychnos Ring is in development, and Alkyon is a concept that flies only in simulation; we write down what we learn in our build log. If the bet is right, the change will not look like a robot butler. It will look like a ring you press, and a small robot that comes when you call, is being taught to find what you have misplaced, and stays out of the way the rest of the time.

Sources, read 4 November 2026: Brohan and colleagues, RT-2: vision-language-action models transfer web knowledge to robotic control, 2023; Gemini Robotics Team, Gemini Robotics: bringing AI into the physical world, 2025 (both arXiv); Gasteiger and colleagues, friends from the future: a scoping review of research into robots and computer agents to combat loneliness in older people, Clinical Interventions in Aging, 2021.