A robot’s training begins before a model sees its first demonstration. It begins with the task someone chooses, the objects they arrange, the cameras they position, and the way they perform the work.
Collecting an episode is easy. Collecting good episodes, ones that push the hardware to its limits with smooth, clean trajectories and no mistakes, is hard. Combining those episodes into a complete dataset — the collective knowledge that gives models not just the ability to repeat, but to generalize — is even harder, because it requires balancing diversity with dataset size, introducing variation without introducing chaos, and having a deep understanding of what exactly happens when a model learns.
That’s the work behind PrismaX. Today, we’re opening Robotics Data in limited release. It brings access to a growing data bank together with custom collection for teams building physical AI.
The release connects the people generating and reviewing demonstrations with the teams using them to train models. It’s the next step in the service layer we’ve been building.
What a demonstration carries
Consider a task as familiar as putting objects into a container. The instruction is simple, but the learning process is complex. Think about how the model learns when to close the gripper. Is it the wrist camera showing the gripper around the target? The arms reaching a certain height above the table? The top camera losing sight of the target?
How the learning happens matters, because it determines how the model behaves in new environments. You might want the model to learn to grasp parts via color cues, or learn the shape, but not the details, of a tool. You might be content to deploy the model in an environment with clean white walls, or you may expect the model to run well under changing light conditions.
The data, more than model size or architecture, determines these behaviors. How it’s selected shapes how much data you need, how large a backbone must be to absorb it, and how much compute training takes.
This is why we think about data quality throughout collection. The hardware setup, scenario design, operator behavior, and review process all influence what ends up in training. Each needs deliberate attention.
Teleoperation captures skilled operators guiding robot arms and humanoids through physical tasks. Egocentric capture records hands-on work from the perspective of the person performing it. They offer different views of interaction, with different uses in training physical AI systems.
Across both, the useful question is specific: what does a team need its model to learn, and what experience would help it learn that?
Answering that question is central to how we work with teams developing world models, vision-language-action models, and other physical AI systems.
The work behind the data
In March, we described the cycle at the center of PrismaX: robots interact with the world, those interactions generate data, and that data helps improve intelligence. As capabilities develop, the next round of interaction can become richer.
Making that cycle work requires coordination. Someone has to set up the hardware, design the tasks, run the sessions, review the results, and learn from what happens during training.
VLA Foundry provides the software through which our network of owner-operators and reviewers contributes to that process. Operators capture demonstrations. Reviewers assess their quality and help identify problems in the collection process. Their work gives the data standard a practical role in everyday operation.
We also train on the data we collect. That gives us another way to examine choices that can be difficult to judge from a recording alone.
In our recent fine-tuning work, for example, changing camera placement in a simulated evaluation affected a model’s ability to place an object accurately. A successful grasp didn’t necessarily mean the rest of the task would succeed. Understanding how the collection setup relates to the evaluation setup mattered.
Findings like these feed back into our collection standards. They help us decide what to document, what to control, and what to investigate next. The relationship between collection and training is part of the service itself.
A data source that develops with the work
A team’s data needs change as its models develop. Early experiments may call for an existing set of demonstrations. Later work may expose a missing task, an unfamiliar environment, or a hardware requirement that needs a different collection setup.
Robotics Data gives teams a way to work with us across those stages.
Discover and Researcher provide subscription access to our growing data bank, with clear pricing and access defined by each plan. Teams can explore the catalog, preview demonstrations, and request access to the plan that fits their work. As we continue collecting, the bank develops over time.
For requirements beyond the catalog, Enterprise supports customized collection. Teams can describe the tasks, embodiments, and scenes they need, and work with us to scope production around them.
This brings teleoperation data, egocentric data, and custom collection into one broader offering. We’re continuing to build its coverage, guided by the needs of the teams using it.
Our role is hands-on. We work through what makes a demonstration useful for a particular training objective, and how to produce it consistently. That connects the data a team receives to the people and processes behind it.
Building with the first cohort
We’re starting with a limited release so we can work directly with the first teams using Robotics Data. We want to understand what they’re training, where the data fits, and what they need next.
Those conversations will help shape the catalog and the product as both develop. Alongside our own training work, they’ll give us more opportunities to examine how the data performs in practice and share what we learn.
In practice, that direction has two parts: Robotics Data for the teams building physical AI, and VLA Foundry for the network capturing and reviewing the experience those teams learn from.
Physical intelligence develops through repeated interaction with the world. We’re building the service layer that helps turn those interactions into useful training data, and brings the lessons from training back into the next round of collection.
Robotics Data is now in limited release. Explore the data and request access today.
