Unprecedented Scale for Physical AI Data Infrastructure
The world's largest egocentric video dataset at factory scale.
Multimodal human interaction data and pre-trained foundation models for robotics, world models, and embodied AI — captured across thousands of hours in real kitchens, factories, dark stores, and construction sites.
Facility 01
Partners & Clients
Partnered with
major robotics labs
training the next
generation.
University research labs and leading robotics companies all over the world trust our data to train their world models for the next generation of home and construction robots.
University Research Labs
Robotics & AI Companies
“We replaced 14 months of internal collection with one Dimag AI Labs contract. The labelling alone would have cost us a year.”
“The temporal sync is the cleanest we’ve worked with. Our world-model team stopped writing alignment patches the day we switched.”
The Facility
A gigantic floor,
purpose-built
for scale.
We don't crowdsource. We staff our own purpose-built data factory with 50+ trained operators and 200+ active headsets running 16 hours a day, every day. When you need 10,000 hours of a specific embodiment, we ship it in weeks — not years.
// 002 — Thesis
You don't get to
pick quality
or quantity.
You need both.
- · Every recording above 1080p, most at 4K stereo
- · Hardware-synced multi-camera + IMU streams
- · Frame-accurate temporal alignment (≤2ms drift)
- · Triple-pass human QA on every clip
- · Fisheye, equirectangular, and custom rigs
- · 200M+ images, 150K+ audio hours indexed
- · 1,140+ distinct scene categories
- · 200+ headsets across multiple facilities
- · Custom collections in 4–6 weeks
- · Indexed, searchable, ready to train
// 003 — Egocentric Samples
Real hands.
Real environments.
A peek into the marketplace. Click any sample to open its Overview, Samples (drive links), and Metadata.
// 004 — What ships with every dataset
Not just video.
A complete
world signal.
Every recording leaves the floor with 14 synchronized modalities — ready for your foundation model, world model, or policy training pipeline.
- Object physics data
- Camera intrinsics
- Depth (stereo + ToF)
- Mesh reconstruction
- Environment maps
- Hand pose (26-keypoint)
- Body pose & mocap
- Eye gaze + saccades
- Pose estimation + tracking
- Action understanding
- Frame-level annotation
- Semantic labelling
- Temporal action segments
- Narrative extraction
- Triple-pass QA
- Timestamps · ns precision
- Location & site context
- Camera shifts & rig calib
- Video metadata schema
- Segmentation masks
// 005 — Capture Fleet
Every embodiment
you'll ever need to
train against.
We run a multi-vendor fleet so your dataset isn't locked to one sensor stack. Aria for stereo SLAM, Ray-Bans for in-the- wild capture, GoPros for action-dense kitchens, and custom fisheye + UMI rigs for robot-equivalent collection.
Project Aria
Stereo SLAM · 7 cameras · 1080p+
Long-form egocentric in domestic + retail scenes
Ray-Ban Meta
Discreet form factor · 1080p · audio
In-the-wild capture, narrative voice memos
GoPro Hero 12
5.3K · HyperSmooth · linear + wide
Construction, kitchens, action-dense scenes
Custom Fisheye Rig
220° FOV · stereo depth · 4K
Wide-context world-model training
UMI Handheld
Manipulation gripper · proprio · RGB
Robot-equivalent embodiment data
Quest 3 + Trackers
6DoF body · hand tracking · passthrough
Mocap, full-body action sequences
// 006 — Coverage
From diverse kitchens
to dark stores, we've
been there.
// 007 — The Loop
The full world-model
production cycle.
We don't drop a dataset and disappear. From the first capture on the factory floor to a deployed policy refining itself on the edge — we own the entire data stack so your team can own the model.
Capture
200+ HMDs run on the factory floor and in-the-wild, generating 3,200+ hours of synced multimodal data each week.
Annotate
Triple-pass labelling, narrative extraction, pose, action, and segmentation — every modality frame-aligned.
Train
Curated splits feed your foundation model, world model, or VLA policy. We ship with reproducible eval suites.
Deploy
Pre-trained checkpoints, embodiment-mapped, ready for RViz, Gazebo, ROS2, and NVIDIA Isaac Sim.
Refine
Edge telemetry flows back into custom collections. The loop tightens with every cycle.
// 009 — Get In Touch
Tell us what
you're training.
Whether you need a slice of the marketplace or a fully custom collection across multiple embodiments — we'll scope it, schedule the factory, and ship in weeks.
- Email founders
- founders@trybibby.com
- WhatsApp / Text
- +1 203 390 8652
- Offices
- Bengaluru · Mumbai
- Data Factory
- Gujarat · India
- Response
- Inquiries answered within 24 hours.
Custom collections quoted in 2 days.
Egocentric robotics data at factory scale
Dimag AI Labs builds physical AI data infrastructure for teams training humanoids, world models, and vision-language-action (VLA) policies. Our egocentric video datasets combine 100,000+ hours of factory-captured footage with stereo depth, 6-axis IMU, hand pose, body pose, gaze, and frame-accurate action labels — indexed across kitchens, construction sites, dark stores, garment lines, and domestic environments.
Unlike crowdsourced collections, we operate dedicated capture facilities in Gujarat, India with 200+ headsets and 50+ trained operators. Every clip passes triple-pass QA and ships with 14 synchronized modalities for foundation model and robotics training pipelines.
Explore our pre-trained robotics models for navigation, grasping, and locomotion — or request a custom collection scoped to your embodiment in as little as 4–6 weeks.