Robotics has always had to cross a boundary that software can often ignore: the world is not finished when the code is. A robot can look capable in a controlled setting and still meet a long tail of objects, surfaces, people, and situations that were never in the room where it learned. The next step is not only to make a stronger model. It is to build a way for the model to keep meeting the world.
We are building around three connected stages: entering the world, learning from it, and deploying back into it. LoopSkill, Looper, and LoopOS are the parts of that loop.
01 / LoopSkill Enter → Scalable Data Collection. The loop starts by entering real-world scenarios and collecting interaction data. For wheeled robots with grippers, this path is already well established. For humanoid robots with dexterous hands, scalable data collection remains a major bottleneck. We see this as a gateway to making robots a more general-purpose physical API.
02 / Looper Pretrain → Steerable Foundation Model. We aim to build an embodied foundation model that generalizes across tasks and can be steered by both humans and agents. Steering should let the model adapt to a specific scenario without changing its parameters. Experience from deployment, across diverse tasks and mixed-quality rollouts, should then feed the next model update.
03 / LoopOS Deploy → Agentic Self-Improvement. Deployment closes the loop. Through agent steering and minimal human supervision, a robot should adapt to new scenarios and improve through real work. The experience it gathers should update both the foundation model and the agent, so that every deployment helps the system adapt more quickly to the next.
We will share this work in small, concrete steps: a report when the work is ready to be read, a demo when the system is ready to be seen, and a note when an idea is ready to be tested in public.