What we’re building
We are building models that turn robot intelligence into precise, reliable physical work, and carry that skill from one embodiment and controller to the next.
- Cross-embodimentOne model across arms, hands, and tools. A skill learned on one embodiment carries to the next.
- Cross-controllerOn a compliant arm, it presses in and searches until a plug finds the socket. On a stiff arm, it aligns first and inserts carefully. Our model adapts its strategy to the controller, without retraining.
- Hybrid actionsOur model chooses how to act at each moment: fast waypoint motion through open space, dense force-regulated control at contact, and the stiffness and force targets each phase needs.
- Closed loopAn interaction policy runs beneath the model and reacts to contact as it happens.
- Force-awareOur model reasons about force as well as position. It decides when the robot should give way and when it should push, and how hard, on whatever controller it runs.
Where we work
Anthropic’s recent study of what work robots can do sorts tasks by how controlled the environment has to be. Some work robots cannot do at all (E0). Some they can do only in cells built around them, like a welding line (E1). Some they can do in structured facilities built for people, like a warehouse (E2), and a sliver anywhere (E3). The study found that half of physical tasks won’t be automated at scale until robots get better at touching and handling objects.
Much skilled contact work sits at the bottom of that ladder. Robots either can’t do it, or do it only in cells engineered around a single part. Even where robots weld, people still position the parts, grind, and finish, and that surrounding work keeps the whole job expensive to automate.
Moving it up the ladder means beating the tradeoff robot automation has always faced: precision against scale. Classical automation is precise and reliable, but every cell is engineered by hand for one task. Generalist models handle many tasks and settings, but stop short of the precision skilled work needs.
We work on both sides of that tradeoff. Human-in-the-loop reinforcement learning on each robot, with every attempt graded against the spec of the work, makes contact work precise and reliable enough to do at all: E0 to E1. Human data at scale, recorded with handheld UMI-style tools, and models that adapt to any robot’s controller take it out of purpose-built cells and into the plants, depots, and data centers where the work actually happens: E1 to E2.
Where we start
We are starting with fastening in the skilled trades, because it is the hardest test of control we know. A single lug nut needs three kinds of it: a fast move to the next nut, a compliant touch to seat the socket, and a firm hold while the impact driver runs. If a model can do that on any robot, it can do the rest of contact-rich work.