PHYSICAL AI RESEARCH LAB

We help train frontier robot models
for the real world.

Our proprietary camera creates the unique real-world datasets these models learn from.

PUBLISHED RESEARCH
More papers each quarter  ·  CVPR  ·  IROS  ·  Humanoids Summit
RESULTS

Published results on the models the field is building on.

We benchmark our data on leading public foundation and robot models. The gains are measured and reproducible.

SABER · Published May 2026

2.19× improvement fine-tuning GR00T N1.6 on dual-stream egocentric data (retail manipulation). SABER · 44.8K samples · 91% fridge success.

2.19X
Improvement
44.8K
Samples
91%
Fridge Success
PRISM · Published March 2026

66.6% error reduction fine-tuning Cosmos-Reason2-2B to state-of-the-art on physical AI benchmarks. PRISM · 270K samples · 5× reasoning.

66.6%
Error reduction
270K
Samples
Reasoning
THE CAPTURE PLATFORM

Data you can't get anywhere else — because we built the camera.

Our results come from a capture system nobody else has: single-shot omnidirectional 3D optics, eight years in production, protected by 32+ patents. The moat is the optics; every dataset inherits it.

Dual-stream capture

Synchronized 360° exocentric + egocentric in a single pass.

Single-shot 360° stereo

Optical design published at CVPR 2016.

8 years in production

Real-world failure modes already engineered out.

32+ patents

A decade of protected optical IP; the platform is the moat.

Explore the capture platform
✦ SCARCE APEX
Real-World 3D Capture
The training signal robots actually need
Simulation
Scalable, but bounded by real-world fidelity
Web Video
Abundant, flat, no 3D structure
Sim quality is bounded by the real data feeding it. We widen the apex.
OUR POINT OF VIEW

Real-world 3D data is the scarce apex of physical AI.

Today’s models train on internet 2D video, simulation, or narrow lab capture — each leaves a gap. Simulation quality is bounded by the real data feeding it. Our omnidirectional 3D platform captures the layer the others can’t reach, and widens the apex.

The same gap caps world models — their realism is bounded by the real-world data behind them.

REAL-TO-SIM-TO-REAL

One pipeline — from real-world capture to VLA/VLM-ready data.

1

Capture

Synchronized Alia 360° + GoPro dual-stream in real environments

2

Annotate

AI-assisted (SAM2, Grounding DINO) + human QA — 10× faster

3

Reconstruct

3D Gaussian Splatting creates photorealistic scenes

4

Convert (USD)

Automated USD export with physics for any USD-compatible simulator

5

Simulate

1,000+ frames/hour with domain randomization

6

Train

Ready for VLA and VLM fine-tuning

🟢 OpenUSD (Native, physics-ready)
🤗 Hugging Face LeRobot (RLDS)
📦 Open X-Embodiment
See the full pipeline
HOW WE PARTNER

Ways to work with us.

For teams building VLA, VLM, and world models.

🤝

Partner on a benchmark

Co-develop and measure gains on your model.

Get in touch →

License a dataset

PRISM (VLM) and SABER (VLA).

Explore datasets →

Simulation assets

Photorealistic USD environments and product assets.

Explore simulation assets →

Custom capture

Our team to your facility; the data stays yours.

Talk to our team →
FAQ

Common questions

What is DreamVu?

DreamVu is a physical AI research lab whose proprietary camera creates the unique real-world datasets that train frontier robot models.

What datasets does DreamVu provide?

PRISM, a vision-language dataset, and SABER, a vision-language-action dataset, both captured with DreamVu's omnidirectional 3D camera platform.

How is DreamVu's data captured?

With a proprietary single-shot omnidirectional 3D camera, protected by 32+ patents and hardened over eight years in production.

Who is DreamVu's data for?

Teams building robot foundation models (VLA), vision-language models (VLM), and world models — anyone whose system has to understand or act in the physical world.

GET STARTED

Building physical AI? Let's partner.

Whether you're training humanoids, building world models, or shipping enterprise robotics — we'll lift your real-world outcomes.