Training Data for 1X Technologies

1X Technologies has been more explicit than any other humanoid company about the centrality of data to their approach. VP of AI Eric Jang — formerly a research scientist at Google Brain where he worked on robotic grasping and reinforcement learning — has publicly argued that the path to android intelligence is fundamentally a data problem. In his widely-read essay 'All You Need Is Data,' Jang argued that given sufficient demonstrations of human behavior, neural networks can learn the visuomotor policies needed for general-purpose robots. This philosophy drives their massive teleoperation infrastructure and creates an insatiable demand for diverse, high-quality demonstration data.

The technical architecture behind 1X's approach relies on two key components. The World Model is a learned simulator that predicts future visual states from current observations and proposed actions. Rather than hand-building physics engines to predict how the world behaves, 1X trains neural networks on video data to learn these predictions implicitly. This approach requires massive quantities of video showing physical interactions in diverse environments — the exact kind of data that purpose-built collection campaigns produce more efficiently than web scraping. Redwood AI is a vision-language model that processes camera feeds and natural language instructions to generate robot control signals in real time. Together, these systems allow NEO to understand verbal commands, visually perceive its environment, predict the consequences of actions, and execute physical tasks.

The end-to-end imitation learning pipeline maps directly from visual observations to motor commands without decomposing robot behavior into separate perception, planning, and control modules. This architectural choice means the network must implicitly learn physics, object properties, spatial relationships, and task structure entirely from raw demonstrations — demanding orders of magnitude more training data than modular systems where each component can be trained on different data types.

NEO's target domain — consumer household assistance — is the most data-hungry deployment context imaginable. Homes vary enormously in layout, furniture, object placement, lighting, and cultural norms. A robot that can fold laundry in one apartment needs demonstrations across different clothing types, folding surfaces, laundry room configurations, and home environments to generalize reliably. The long tail of household variation — regional appliance differences, cultural norms around food preparation, idiosyncratic organization systems — is essentially infinite. No amount of data from a single teleoperation studio can cover this variation.

The consumer price point creates additional pressure. At $20,000, NEO must work reliably out of the box in each customer's unique home environment. Unlike industrial robots that can be professionally calibrated for a specific workspace, consumer robots face immediate deployment in unknown environments. This requires either exhaustive pretraining data that covers the full distribution of home environments, or extremely efficient adaptation mechanisms that can customize behavior with minimal local data. 1X is betting on the former approach — and that bet requires data at unprecedented scale and diversity.

As 1X scales from hundreds to thousands of deployed NEO units, each robot operating in a unique home becomes both a deployment endpoint and a potential data collection node. But bootstrapping this fleet requires initial training data from diverse homes before deployment — a chicken-and-egg problem that external data collection partnerships solve directly.