The DreamVu Pipeline
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 research in this field, so we know the difference, 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.
Stage 01
Capture
Exocentric
Alia, our own camera. A single-shot 360° stereo panorama with depth, from one camera position. The whole room, not a chosen angle.
Egocentric
A head-mounted camera on the worker, synchronized to the Alia stream frame for frame.
Wrist
An optional third camera for close manipulation and grasp detail.
Other cameras
If a program needs cameras other than ours, we use those too. The pipeline does not depend on any one camera.
Environments
Working venues during working hours. We have access to hundreds of venues across a range of industry segments.
Stage 02
Annotate
Method
AI does the first pass. A person reviews every batch. Segmentation, tracking, action boundaries, and task breakdown are set per program.
Components
The best current model for each task, replaced as the field moves. Our own components where published methods are not good enough.
Capacity
20,000 hours a month of fully annotated data.
Stage 03
Deliver
| Output | Used for |
|---|---|
| VLM training datasets | Vision-language model fine-tuning |
| VLA training datasets | Manipulation and action policy training |
| Simulation-ready USD assets | Any USD-compatible simulator |
| High-resolution walkthrough video | World model and generative video training |
| Custom formats | Whatever you specify |
Standards
OpenUSD. Hugging Face LeRobot (RLDS). Open X-Embodiment.
If you need something that is not on this list
Tell us and we will scope it. We also run 3D Gaussian Splatting reconstruction, USD conversion with physics, and domain-randomized rendering. We run whatever your model needs.
Quality
Four gatesEvery capture passes four gates before we deliver it.
G1 Calibration
Checked for every rig, every session.
G2 Alignment
All camera streams confirmed in sync before annotation starts.
G3 Annotation review
A person reviews every batch.
G4 Rejection
Batches that fall below the threshold are captured again.
Full numeric specification available under NDA
Tell us what your model needs to learn.
Capture programs, research collaboration, and dataset partnerships.