Real-world data for training robot foundation models and world models.
DreamVu is an applied research company. We built the camera, we publish the research, and we run the operation that captures real work in real places.
What we deliver.
Every clip below was captured in a real working environment. Nothing is staged and nothing is synthetic.
What you can get, from one capture system.
What we capture depends on what your model needs to learn. The camera configuration and the annotation change; the venue access, the operation, and the quality standard do not.
| Output | Capture | For |
|---|---|---|
| Robot training data | Manipulation, VLM, and navigation datasets. Ego, exo, and wrist cameras as the dataset needs. Alia, or cameras you specify. | Training VLA and VLM models |
| World model data | Alia 16K. A 360° stereo panorama with depth for every pixel. | Training generative video and world models |
| Simulation assets | Objects scanned from real venues. USD with physics. Environments in development. | Training robots in simulation |
The same venue access supports all three. Capture resolution is set by what the output needs — 16K for world model data, lower for robot training, where 16K is more than a model can use. Object libraries are built through a separate process.
For robot training data we will use whatever cameras a program calls for, including your own. World model data needs the 360° stereo and depth that only Alia captures. Navigation datasets are ego captures and need no Alia at all.
We customize the pipeline for every data challenge.
Each stage uses the best method available — open source where open source is better, our own where it is not. We publish in this field, so we track what changes, and we swap components when something better comes out. You tell us what your model needs to learn, and we build the pipeline for it.
Capture
Alia is our own camera and our first choice. If a program needs other cameras, we use those too.
Annotate
AI does the first pass. A person reviews every batch, at production volume.
Deliver
Robot training data, world model data, or simulation assets, in the format your training stack uses.
An exocentric view with real 3D geometry.
Alia has multiple sensors in one housing, angled to cover the full sphere. Every sensor fires at the same instant and our software combines them into one 360° panorama with depth for every pixel. Rigs built from separate cameras have to be synchronized, and they drift. Below, both outputs from the same capture.
The same instant, two ways. Color at 16K, metric depth at every pixel, across the full sphere.
Capacity and access, already built.
Getting a capture team into a working pharmacy, an auto plant, or a hospital ward takes agreements, training, and compliance approvals. We have already done that.
Examples of where we capture
We publish what we learn.
We test our capture and annotation methods on public models and publish the results. Four papers so far, alongside our granted patents and the earlier computer vision work our team brought to the company.
Measured on public foundation and robot models. Published, with reproducible results.
Common questions.
What does DreamVu do?
What can I buy?
How is your data different?
Do you sell cameras?
Who is this for?
Tell us what your model needs to learn.
Capture programs, research collaboration, and dataset partnerships.