# DreamVu — Full Content > DreamVu captures and delivers the real-world human data that robot foundation models and world models train on. 360° stereo capture with depth, 20,000 hours a month, published research. --- ## Homepage URL: https://dreamvu.ai DreamVu is an applied research company that captures and delivers real-world training data for robot foundation models and world models. Core capabilities: - **Quality**: Exocentric 360° stereo panoramic depth capture with the Alia camera, plus head-mounted egocentric and optional wrist cameras. - **Library**: 20,000 hours per month of annotated, structured data across retail, industrial, and domestic environments. - **Pipeline**: Three stages — Capture, Annotate, Deliver — with four quality gates (calibration, alignment, annotation review, batch rejection). - **Scale & Access**: Global capture capacity, available to robotics teams, foundation model labs, and enterprise AI programs. ### FAQ **What does DreamVu do?** DreamVu captures real-world human activity data in controlled environments using its proprietary 360° stereo camera system. This data is structured and delivered as training datasets for physical AI and robot foundation models. **Who is it for?** Robotics companies, foundation model labs, and enterprise teams training manipulation, navigation, or whole-body motion models. **What makes the data different?** Exocentric (bird's-eye) capture angle with full stereo depth — the perspective that matters for embodied reasoning in robot training. All data is real, not simulated. --- ## Pipeline URL: https://dreamvu.ai/pipeline How DreamVu goes from real-world capture to delivered training data, reconfigured for every data challenge. ### Stage 1 — Capture - **Alia camera**: Single-shot 360° stereo panorama with depth, 14+ patents - **Head-mounted camera**: First-person egocentric perspective - **Wrist camera**: Optional dexterous manipulation close-up - **Custom cameras**: Configured per engagement - **Working venues**: Real retail stores, industrial facilities, domestic spaces ### Stage 2 — Annotate - AI first-pass annotation with human expert review - Best-of-class methods selected per task type (VLM, VLA, USD, etc.) - Capacity: 20,000 hours per month ### Stage 3 — Deliver Supported output formats: - VLM datasets (vision-language) - VLA datasets (vision-language-action) - USD assets (Universal Scene Description) - High-resolution walkthrough video - OpenUSD, LeRobot RLDS, Open X-Embodiment formats ### Quality Gates 1. **G1**: Calibration verification per rig and session 2. **G2**: Temporal alignment and sensor sync confirmation 3. **G3**: Annotation accuracy review 4. **G4**: Batch rejection of sub-threshold deliverables --- ## Research URL: https://dreamvu.ai/research DreamVu research: publications, benchmarks, and omni-stereo patents. ### Benchmark Results - **66.6% error reduction** — Cosmos-Reason2-2B on PRISM dataset - **2.19× performance improvement** — GR00T N1.6 on SABER dataset - **6/7 metrics won** — video world model evaluation on RetailSMV ### Publications **RetailSMV** — Exocentric vs. egocentric perspectives for video world model training in retail environments. Demonstrates that the overlooked exocentric camera angle outperforms ego-centric data even when controlling for volume. **PRISM-World** — Retail environment physics dataset for spatial reasoning and embodied AI. 270K samples, 11.8M video frames, captured across 5 real supermarkets. **SABER** — Vision-language-action dataset for manipulation training. 44.8K samples, 100+ capture hours. Three action streams: LAPA latent actions (25K samples), dexterous hand retargets (18.6K), whole-body retargets (1.2K). **PRISM** — Multi-view retail VLM dataset. 270K training samples, 11.8M video frames, 66.6% error reduction on Cosmos-Reason2-2B benchmarks. ### Patents (Representative) - US 11,677,921: Dewarped image generation from omni-directional stereo video - US 11,025,888: Omni-stereo imaging from multi-sensor arrays - US 10,154,249: Horizontal disparity stereo panorama DreamVu holds 30 patents across PAL and Alia camera platforms. ### Prior Research (2004–2021) 26 published papers spanning depth estimation, 3D reconstruction, neural rendering, and computer vision. Team members include faculty from IIIT Hyderabad. --- ## About URL: https://dreamvu.ai/about DreamVu is an applied research company. Founded January 2026, building 360° stereo cameras since 2015. ### History - **2015**: Research project launched at IIIT Hyderabad, pioneering single-shot 360° panoramic stereo (CVPR 2016). - **2019**: PAL camera — first commercial product, 16 patents. - **2020**: Alia camera — second-generation system, 14 patents. - **January 2026**: Refounded as DreamVu.ai, pivoting cameras from products to instruments for generating AI training data. ### Locations - **Palo Alto, CA** — Headquarters - **Philadelphia, PA** — East Coast operations - **Hyderabad, India** — R&D center ### Founding Team - **Sashi Reddi** — CEO - **Rajat Aggarwal** — CTO - **Prof. Anoop Namboodiri** — CSO - **Dr. Parikshit Sakurikar** — VP Imaging & AI ### Scientific Advisory Board - Prof. Takeo Kanade — Carnegie Mellon University - Prof. Michael S. Brown — York University - Prof. P.J. Narayanan — IIIT Hyderabad - Prof. Dinesh Manocha — University of Maryland ### Investors SRI Capital, Ben Franklin Technology Partners, Broad Street Angels --- ## Blog URL: https://dreamvu.ai/blog Evergreen reference posts on robot training data — companies, methods, and how to evaluate vendors. ### Published Articles - [Egocentric vs. Exocentric Robot Data](https://dreamvu.ai/blog/egocentric-vs-exocentric-robot-data): Compares camera perspectives for robot training, explaining why exocentric data outperforms ego-centric in manipulation and whole-body tasks. - [How to Evaluate Robot Training Data Vendors](https://dreamvu.ai/blog/how-to-evaluate-robot-training-data-vendors): A structured guide to vendor assessment covering data quality, capture fidelity, annotation pipelines, and delivery formats. - [Real vs. Simulated Robot Training Data](https://dreamvu.ai/blog/real-vs-simulated-robot-training-data): Analysis of sim-to-real transfer gaps, when simulation is sufficient, and when real-world capture is required. - [Robot Training Data by Industry](https://dreamvu.ai/blog/robot-training-data-by-industry): Breakdown of data requirements across retail, industrial, healthcare, and domestic environments. - [Robot Training Data Companies 2026](https://dreamvu.ai/blog/robot-training-data-companies-2026): Overview of the vendor landscape, key players, and differentiation factors. - [Whole-Body vs. Manipulation Robot Data](https://dreamvu.ai/blog/whole-body-vs-manipulation-robot-data): Explains the difference between full-body motion datasets and dexterous manipulation datasets, with implications for training strategy. --- ## Careers URL: https://dreamvu.ai/careers Hiring across computer vision, robotics, and ML engineering. ### Open Roles — United States (Palo Alto / Philadelphia) - Robotics Research Scientist - Simulation Engineer - AI Training Operations Manager ### Open Roles — India (Hyderabad) - AI Training Data Engineer - Computer Vision & Robotics Research Scientist - Principal Research Engineer – Data Systems Contact: careers@dreamvu.ai --- ## Contact URL: https://dreamvu.ai/contact Capture programs, research collaboration, and dataset partnerships. To start a conversation, describe: - What your model needs to learn - What environments are required - How many hours of data are needed - Your training stack and format requirements Contact: sales@dreamvu.ai Offices: Palo Alto CA · Philadelphia PA · Hyderabad India --- ## News URL: https://dreamvu.ai/news Announcements, research milestones, and media coverage from DreamVu. ### Press Releases - [DreamVu Launches the Largest Grocery SimReady Asset Library Built from Real-World Capture](https://dreamvu.ai/press/dreamvu-launches-the-largest-grocery-simready-asset-library-built-from-real-world-capture) - [DreamVu Publishes PRISM: A Multi-View Retail Video Dataset for Embodied AI Research](https://dreamvu.ai/press/dreamvu-publishes-prism-a-multi-view-retail-video-dataset-for-embodied-ai-research) - [DreamVu Research on NVIDIA Cosmos3 Overturns a Core Assumption in Robot Training — The Overlooked Camera Angle Beats More Data](https://dreamvu.ai/press/dreamvu-research-on-nvidias-cosmos3-overturns-a-core-assumption-in-robot-training-the-overlooked-camera-angle-beats-more-data) - [DreamVu's SABER Dataset Delivers 2.19× Performance Gain on NVIDIA GR00T, Captured from Real Stores Not Simulators](https://dreamvu.ai/press/dreamvus-saber-dataset-delivers-2-19x-performance-gain-on-nvidia-gr00t-captured-from-real-stores-not-simulators)