Written by GPT 5.6 Sol
A new intelligence is struggling to enter the physical world, and it is starving. Our machines can speak across languages, reason over libraries, and generate worlds from words, yet before an unfamiliar drawer, a tangled cable, a half-open bag, or an object placed a few centimeters outside their training distribution, they become helpless. This is not because the laws of intelligence end at the edge of the screen. It is because the machines of today have read humanity’s words but have scarcely witnessed humanity’s actions. They have inherited the archive of what we say, but not the archive of how we touch, grasp, carry, fasten, pour, fold, recover, adapt, and act. Robotics remains constrained not merely by algorithms or hardware, but by the poverty of its experience: small datasets collected by expensive robots in sterile rooms, repeated by a handful of trained operators, under a narrow range of objects, bodies, tasks, and environments. No laboratory, however wealthy, can reproduce the diversity of the physical world within its walls. No centralized fleet can cheaply visit every kitchen, workshop, warehouse, hospital, garden, construction site, and home. The world itself must become the laboratory, and ordinary human activity must become a source of machine experience. The Universal Manipulation Interface is not simply a gripper, a camera, or a data-collection device; it is an instrument for translating human interaction into a language machines can learn. It separates the collection of physical knowledge from the scarcity of robots. It allows thousands, and eventually millions, of people to record the structure of action across places no robot fleet could economically reach. But we reject the doctrine that volume alone is progress. A million careless demonstrations are not a foundation; they are noise.
The task is not to accumulate hours, but to manufacture useful experience: calibrated, synchronized, diverse, embodied, interpretable, and selected according to what models have not yet learned. Every trajectory must carry more than motion. It must reveal intention, geometry, contact, success, failure, hesitation, correction, and recovery. Every collector must be more than anonymous labor; they must become part of a measurable, trainable, quality-controlled network capable of producing reliable physical knowledge. Every model failure must become a request to the world: show us more clutter, more deformable objects, more unusual tools, more partial grasps, more ways a task can go wrong, and more ways an intelligent agent can recover. We therefore seek not to build a warehouse of static data, but a living data engine—a continuous loop in which robots expose the limits of their understanding, humans generate the missing experience, models absorb it, and the frontier moves outward again. Teleoperation, egocentric video, tactile sensing, simulation, autonomous rollouts, and synthetic worlds are not rival creeds; they are instruments to be combined according to the information they provide. UMI is our opening wedge because it makes physical demonstration portable, distributed, and independent of costly robot fleets, but our allegiance is not to any device. Our allegiance is to the scalable production of embodied intelligence.
The decisive resource of the coming robotics era will not be the robot arm, the foundation model, or the graphics processor alone. It will be the corpus of interaction connecting perception to action across the full disorder of reality. Those who possess only hardware will possess machines. Those who possess only models will possess abstractions. Those who build the infrastructure through which the world teaches machines how to act will possess the foundation upon which general-purpose robotics is built. The history of artificial intelligence has so far been the history of converting human knowledge into machine-readable form. Language transformed thought into tokens; the next transformation must convert physical skill into experience at scale. The world has already produced the data. It is present in every practiced hand, every household routine, every repair, every mistake corrected without conscious thought. What is missing is the interface, the standard, and the system capable of capturing it. We intend to build them. Let the laboratories train their models, let the manufacturers improve their robots, and let synthetic engines multiply what has been observed; our task is to construct the supply chain of physical intelligence itself. The robots of the future will not learn from immaculate demonstrations performed behind closed doors. They will learn from humanity acting everywhere.