DreamVu — The Data Infrastructure for Physical AI

The most influential minds in AI — from Fei-Fei Li to Yann LeCun — agree: the next frontier isn't larger language models. It's machines that understand and interact with the physical world. But world models, spatial intelligence, and humanoid robots all hit the same wall — they need massive amounts of high-fidelity 3D training data from real environments. That data barely exists today.

We're building it.

DreamVu's proprietary dual-stream capture system — Alia 360° omnidirectional camera + GoPro egocentric — produces synchronized ego-exo 3D training data at a scale and fidelity no one else can match. 360° depth + RGB, no blind spots, no stitching — from a single sensor protected by 32+ patents. Born from breakthrough research at IIIT Hyderabad (CVPR 2016), refined through 8+ years of production deployment.

Any environment. Any domain. Any modality. Our capture platform serves the entire Physical AI stack — from world models and VLMs to VLA foundation models and sim-to-real transfer pipelines.

What You'll Find Here

📦 Datasets — We're releasing curated ego-exo datasets to fuel open research, starting with grocery retail — 500+ distinct skills annotated across shopper behavior, product interactions, restocking, shelf monitoring, and backend operations. Think of it as our ImageNet moment for Physical AI: a public benchmark the community can build on, while our full-scale capture platform serves any domain.

🔬 Models — Fine-tuned state-of-the-art models spanning Vision Language Models, Vision-Language-Action (VLA) foundation models, world models, and more — trained on DreamVu's dual-stream data and demonstrating measurable improvements over baselines.

🌐 World Models & Simulation — Datasets and pipelines built for training world models that understand physics, space, and time — with native support for sim-to-real transfer through NVIDIA Isaac Sim.

🚀 Demos — Interactive Spaces showcasing our models on real-world scenarios with chain-of-thought reasoning, spatial understanding, and action planning.

Why We Started with Grocery

A grocery store contains more distinct manipulation tasks per square foot than almost any other environment — picking, placing, stacking, scanning, bagging, mopping, organizing. It's the ideal proving ground to demonstrate what our capture platform can do. If a model can handle a cluttered grocery aisle with customers, carts, and staff in motion, the same approach transfers to warehouses, fulfillment centers, manufacturing, healthcare, and beyond. The grocery dataset is our open contribution to the research community — the platform behind it works everywhere.

By the Numbers

📊 16,000+ hours of enriched 3D video targeted by Dec 2026 · 500+ distinct skills captured · 5 research papers · 32+ patents protecting our capture technology

Platform Compatibility

NVIDIA Isaac Sim (USD) · Hugging Face LeRobot (RLDS) · Open X-Embodiment · NVIDIA Cosmos · NVIDIA GR00T · AGIBOT World Challenge 2026

🌐 dreamvu.ai · 💼 LinkedIn · 💻 GitHub