Data Infrastructure for Physical AI and Humanoids
Omnidirectional 3D capture, enriched annotation, and simulation-ready data to train the next generation of humanoid robots and embodied AI.
PUBLISHED · MAY 2026
SABER
A Scalable Action-Based Embodied Dataset for Real-World VLA Adaptation. A high-fidelity action dataset built from natural human behavior.
- 44.8K samples
- 100+ hours
- 2.19X improvement
PUBLISHED · MARCH 2026
PRISM
Unifying physical AI knowledge across space, physics, and embodied action. Fine-tunes NVIDIA Cosmos-Reason2 to state-of-the-art on physical AI benchmarks.
- 66.6% error reduction
- 270K samples
- 11.8M frames
Real-world 3D data is the scarce apex of Physical AI.
Today’s robots are trained on internet 2D video, simulation, or narrow-field-of-view lab capture. Each leaves a gap. Our omnidirectional 3D platform captures the layer the others can’t reach.
Purpose-built for Physical AI training data.
A proprietary capture system, eight years of production hardening, and an end-to-end pipeline that turns real environments into simulation-ready training data.
Dual-Stream Capture
Synchronized 360° exocentric (Alia) + egocentric (GoPro) in a single pass. Complete spatial context — what the robot sees and how it appears in the scene.
CVPR-Pedigree Optics
Single-shot 360° stereoscopic design from IIIT Hyderabad, peer-reviewed at CVPR 2016. The optical breakthrough behind every dataset we ship.
8 Years in Production
Deployed across AMRs, surveillance, and smart city. Real-world failure modes already engineered out before the first humanoid dataset.
32+ Patents
A decade of protected optical IP. The capture platform is the moat — every downstream product inherits its quality.
One pipeline. Real-world capture to VLA/VLM-ready training data.
We deliver real-world capture, annotation, 3D reconstruction, USD simulation export, and synthetic generation — as one infrastructure layer.
- Capture
- Synchronized Alia 360° + GoPro dual-stream in real environments
- Annotate
- AI-assisted (SAM2, Grounding DINO) + human QA — 10× faster
- Reconstruct
- 3D Gaussian Splatting creates photorealistic scenes
- Convert (USD)
- Automated USD export with physics for any USD-compatible simulator
- Simulate
- 1,000+ frames/hour with domain randomization
- Train
- Ready for VLA and VLM fine-tuning
Four ways to deploy Physical AI faster.
PRISM — VLM Datasets
Vision-Language datasets that unify space, physics, and embodied action.
SABER — VLA Datasets
Vision-Language-Action datasets on dual-stream egocentric data.
Simulation Assets
Photorealistic USD environments and product assets.
Custom Capture
Our team to your facility. Data stays yours.
Validated on leading Physical AI models.
SABER · Published May 2026
A Scalable Action-Based Embodied Dataset for Real-World VLA Adaptation. Fine-tunes NVIDIA GR00T N1.6 on dual-stream egocentric data to achieve 2.19X improvement over baselines in retail manipulation.
- 2.19X Improvement
- 44.8K Samples
- 91% Fridge Success
PRISM · Published March 2026
The first dataset to unify space, physics, and embodied action in a single real-world deployment domain. Fine-tunes NVIDIA Cosmos-Reason2-2B to state-of-the-art on physical AI benchmarks.
- 66.6% Error reduction
- 270K Samples
- 5× Reasoning
Building Physical AI? Let’s talk.
Whether you’re training humanoids, building world models, or deploying enterprise robotics — our data is built for your stack.