What we’re exploring
Robots that learn.
Untitled Robotics is an early-stage studio working on robotics and AI in the physical world. We’re at the stage where the questions are more interesting than the answers — so here are the questions.
Learning from interaction
Most of what a robot needs to know is hard to write down. Reinforcement learning is the obvious lever and also the most frustrating one: reward design turns into a second job, sample efficiency decides what is even possible, and a real robot cannot afford to fall over ten million times to find out.
World models
A robot that can predict what happens next can plan instead of guess. World models — and the action models being learned from video and interaction — are the part of this field moving fastest right now, and the part we keep coming back to. How much of control can be learned as prediction, and where does that stop working?
The gap between simulation and the floor
Everything works in simulation. The interesting engineering starts at the point where it doesn’t transfer — contact, friction, latency, a cable in the wrong place, a floor that isn’t the floor you trained on. Sim-to-real is where most of the honest work lives.
Small experiments, in public
We’d rather run a small experiment this week than plan a large one for next quarter. Most of them won’t work. We intend to write about those too — the failed run is usually the more useful post.