OFFERINGS
Three product families. Plus Custom Capture.
PRISM Vision-Language datasets. SABER Vision-Language-Action datasets. Simulation-ready USD environments. Custom Capture engagements. All built on the same dual-stream capture platform — extending across twelve industries.
PRISM: VLM Datasets
2 datasets · 10 modalities
PRODUCT FAMILY · PRISM
PRISM — Vision-Language Model Datasets
Richly annotated dual-stream datasets for fine-tuning frontier Vision-Language Models on real-world physical AI tasks.
Annotation Depth: 10 Modalities per Frame
PRISM datasets include modalities 1–10 — the rows labeled "ALL" in the stack below. Every dataset ships with this full annotation stack.
1
Raw Omnidirectional Exocentric Capture
Full-sphere Exocentric RGB + depth from Alia 360° sensor
2
Raw Egocentric Capture
RGB + depth from Ego camera(s)
3
3D Reconstruction
Dense neural 3D scene from omnidirectional input
4
Spatial Semantics
Navigable paths, obstacles, zones, and surfaces
5
Object Semantics
Per-object identity, class, pose, and attributes
6
Physics Metadata
Object mass, friction, deformability — sim-transfer-ready
7
Agent Tracking
People, carts, and robots — trajectories over time
8
Skills & Activities
What each agent is doing — pick, place, scan, stack, mop …
9
Temporal Context
Time-of-day, traffic density, seasonal and layout variants
10
Ego-Exo Synchronization
RGB Video Synchronization
Modalities 1–10 — included in all PRISM datasets
PRISM Grocery
Published March 2026
270,000 annotated samples · 11.8M frames · 500 hours dual-stream capture.
66.6% error reduction on physical AI benchmarks; fine-tunes NVIDIA Cosmos-Reason2-2B.
66.6% error reduction on physical AI benchmarks; fine-tunes NVIDIA Cosmos-Reason2-2B.
PRISM releases are planned across all twelve industries on /industries.
SABER: VLA Datasets
2 datasets · 13 modalities · dual-stream
PRODUCT FAMILY · SABER
SABER — Vision-Language-Action Model Datasets
A Scalable Action-Based Embodied Dataset for Real-World VLA Adaptation. The first high-fidelity retail robotics action dataset built from natural human behavior — not teleoperated demos.
Annotation Depth: All 13 Modalities per Frame
SABER datasets include all 13 modalities — the 10 PRISM modalities plus 3 action-specific rows for VLA training.
1
Raw Omnidirectional Exocentric Capture
Full-sphere Exocentric RGB + depth from Alia 360° sensor
2
Raw Egocentric Capture
RGB + depth from Ego camera(s)
3
3D Reconstruction
Dense neural 3D scene from omnidirectional input
4
Spatial Semantics
Navigable paths, obstacles, zones, and surfaces
5
Object Semantics
Per-object identity, class, pose, and attributes
6
Physics Metadata
Object mass, friction, deformability — sim-transfer-ready
7
Agent Tracking
People, carts, and robots — trajectories over time
8
Skills & Activities
What each agent is doing — pick, place, scan, stack, mop …
9
Temporal Context
Time-of-day, traffic density, seasonal and layout variants
10
Ego-Exo Synchronization
RGB Video Synchronization
11
Hand Pose & Trajectory
Frame-level manipulation actions from egocentric view
12
Body Pose & Trajectory
Frame-level manipulation actions from egocentric view
13
Human to Robot Retargeting
Conversion of human trajectories to standard robot joint data
Modalities 1–10 — included in all datasets (PRISM & SABER)
Modalities 11–13 — SABER datasets only (action data)
SABER Grocery
Published May 2026
44.8K samples · 100+ hours · 2.19× improvement over baselines.
91% success rate on the fridge-restocking benchmark.
Fine-tunes NVIDIA GR00T N1.6 on dual-stream egocentric data.
Built from natural human behavior in real retail environments — not teleoperated demos.
91% success rate on the fridge-restocking benchmark.
Fine-tunes NVIDIA GR00T N1.6 on dual-stream egocentric data.
Built from natural human behavior in real retail environments — not teleoperated demos.
SABER releases follow PRISM releases across the twelve industries on /industries.
Simulation Assets
2 asset types — USD format
Simulation-ready grocery environments and objects in Universal Scene Description (USD) format, built from DreamVu's 3D reconstructions. Drop directly into NVIDIA Isaac Sim or Omniverse for robot training, validation, and sim-to-real transfer.
1
Complete USD Store
Environments
5 photorealistic digital
twin grocery stores — full aisle layouts, checkout areas, stockrooms, and deli counters
2
2,000+ Individual Product
Assets
Grocery items across
produce, dairy, packaged goods, and frozen — each with accurate geometry and PBR textures
3
Physics Properties per
Object
Mass, friction
coefficients, deformability, and collision meshes — ready for contact-rich manipulation simulation
4
Shelf Planograms & Aisle
Topology
Real product placement
layouts with navigable aisle graphs — matches actual store configurations
5
Lighting & Layout
Variants
Multiple lighting presets
(daylight, fluorescent, night) and seasonal layout configurations per store
6
SKU & Barcode Metadata
Product identity, category
labels, and barcode data attached to every asset for scan-and-pick training
7
Grasp Point Annotations
Pre-computed grasping
points and approach vectors for each product asset — accelerates manipulation policy training
8
Isaac Sim & Omniverse
Compatible
Native USD/USDA format —
load directly into NVIDIA simulation tools with zero conversion overhead
Custom Capture
On-site data collection & processing
DreamVu deploys to your facility with our Alia 360° capture rigs and full annotation pipeline. You get the same multi-layer annotation stack applied to your specific environment, operations, and use cases.
On-Site Capture
Alia 360° rigs deployed to your location — warehouses, factories, retail,
hospitals, or any operational environment.
Custom Annotation
Full 7- or 9-layer annotation stack tailored to your domain-specific skills,
objects, and workflows.
3D Reconstruction
NuRec pipeline produces dense 3D scenes — point clouds, meshes, and Gaussian
splats from your facility.
USD Digital Twin
Your environment converted to simulation-ready USD assets for Isaac Sim and
Omniverse workflows.
Model Fine-Tuning
Optional foundational model fine-tuning on your custom dataset for
domain-specific world models or robot policies.
Exclusive License
Custom capture data is exclusively yours — never shared, sublicensed, or
added to our catalog without your consent.