01
Robotics & Physical AI data
Human activity and interaction datasets for robotics and embodied AI — from task definition to training-ready episodes.
DemonstrationsEpisodes
Explore Robotics & Physical AI data
Real captureInteractive capture workspace
Pick it up. That’s training data.
Move, pick, rotate and arrange objects to explore how real-world human interactions become training data for AI.
Loading 3D workspace
Recording session · ready
Egocentric data · Robotics · Physical AI
Real-World Data for the Next Generation of AI
From egocentric video capture to structured, annotated datasets, Mvaak helps power robotics, computer vision, and Physical AI with real-world human activity data.
For robotics & Physical AI · computer vision · speech & voice AI · NLP & LLM data
Rec00:02:12.0Ego-cam · 1080p30 · wide-angle
action reach · sheet_stack
Action log demo
Sample metadata
- Task
- Object handling
- Environment
- Metal fabrication plant
- Capture
- Egocentric · wearable
- Source
- 1920×1080 · 30 fps
- Layers
- Action segments
Tap the frame to inspect objects.
Real Mvaak egocentric capture. The action log and object boxes are demonstration labels placed by hand on this footage — not model output or a client deliverable.
Data at a glance
- Data delivered
- 100,000+
- Data delivered
- Partners
- 200+
- Partners
- States covered
- 6 → 18
- States covered
- Live today, with a defined ramp toward pan-India coverage
- Trained collectors onboarded
- 250
- Trained collectors onboarded
- Network designed to scale within 90 days of programme start
- Environment types
- 40+
- Environment types
- Household, commercial and agricultural settings
- Languages
- 12+
- Languages
- For paired audio, narration and instruction capture
Figures provided by Mvaak; capacity position as of August 2026.
Egocentric samples
Real-world tasks, seen from the first person.
Real work in factories, workshops, building sites and homes — recorded from the viewpoint of the person doing it.
- Real capture · video
Manufacturing Metal fabrication plant
Handling metal sheets
Gloved hands lift and position large steel sheets beside a machine on a busy shop floor.
Egocentric · wide-angle
- Real capture · video
Manufacturing Garment factory
Handling fabric at a sewing station
Both hands gather, fold and feed fabric at an industrial sewing machine — deformable-object handling.
Egocentric · wide-angle
- Real capture · video
Textiles & craft Handloom workshop
Weaving on a handloom
Hands working the loom as the striped fabric builds up — repetitive, two-handed, rhythm-driven work.
Egocentric · wide-angle
- Real capture · video
Manufacturing Metal fabrication unit
Machine tending at a spot welder
A co-worker passes a metal panel; gloved hands position it under the welding electrode — precise two-hand placement.
Egocentric · wide-angle
- Real capture · video
Household Home kitchen
Cooking and kitchen work
Reaching across the stove and counter, handling containers and utensils during everyday cooking.
Egocentric · wide-angle
- Real capture · video
Construction Building site · scaffolding
Finishing work on site
Hands working at a concrete column from the scaffold, with tools and materials within reach.
Egocentric · wide-angle
- Real capture · video
Construction Interior fit-out
Prep work at a tiling site
Hands handle a thin line along the wall above an unfinished floor — precise work in a cluttered space.
Egocentric · wide-angle
- Real capture · video
Warehouse & logistics Industrial storage floor
Moving a pallet jack
Steering a pallet jack between stacked crates and containers — locomotion with a tool in hand.
Egocentric · wide-angle
Every clip is an 8-second excerpt of real first-person capture, shown muted and unannotated. Full episodes and labelled samples are shared on request.
Request a sampleHow it is captured
Four ways we record the real world.
The rig is chosen per task. Every stream is calibrated and aligned to one timeline before it reaches you.
Egocentric / first-person
What the person sees while doing the task — hands, objects and the order of actions, from the robot’s own point of view.
- Head- or chest-level camera
- Action camera, smart glasses, XR headset or research glasses
- Resolution and frame rate agreed per task
Standard deliverable
Stereo video
Two calibrated views of the same scene, so depth and 3D position can be recovered for spatial understanding and manipulation.
- Calibrated stereo pair
- Per-rig calibration results
- Depth or 3D hand pose where the rig supports it
Standard deliverable (RGB or stereo)
Mono-RGB video
Standard colour video of everyday activities — the base input for detection, tracking and activity recognition models.
- Calibrated RGB
- First-person, or paired fixed third-person views
- Temporal segmentation to sub-action
Standard deliverable; paired views optional
Wearable & sensor data
Motion signals that tell a model how the body and head moved between frames, aligned to the video timeline.
- IMU clock-aligned to video frames
- Synchronisation check per rig
- Optional gaze, depth or narration tracks
Standard deliverable; extras depend on the rig
Supported environments
Collected where the work actually happens.
Real homes, shops, kitchens, warehouses and farms — with site permissions and consent arranged before recording.
Households & residential
Everyday manipulation in real homes.
- Cooking and prep
- Opening a drawer
- Folding laundry
In Mvaak’s scope
Retail stores
Shop floors and stockrooms.
- Handling products
- Restocking shelves
In Mvaak’s scope
Restaurants & commercial kitchens
Food-service lines and prep areas.
- Preparing a workspace
- Sorting items
In Mvaak’s scope
Warehouses & logistics
Picking, packing and moving items at volume.
- Picking
- Sorting
- Packing
In Mvaak’s scope
Agricultural settings
Manual field and post-harvest work.
- Transplanting
- Inspecting produce
- Sorting and grading
In Mvaak’s scope
Offices & commercial spaces
Shared workspaces and front-of-house areas.
- Organising items
- Item handling
Confirmed during scoping
Industrial workplaces
Operational settings with equipment and multi-step procedures.
- Equipment interaction
- Tool handling
Confirmed during scoping
Project-specific sites
Other locations, agreed with you and the site owner.
- Defined in your capture brief
Confirmed during scoping
Services
Services built for model teams.
From first-person capture to reviewed, training-ready datasets.
- Real capture
- 02
Egocentric video collection
First-person video of real tasks, captured on agreed rigs with synchronised sensor data.
RGB · stereoIMU
Explore Egocentric video collection
Real capture - 03
Computer vision data
Video labelled for detection, tracking, activity and temporal segmentation.
BoxesTracksSegments
Explore Computer vision data
Illustrative frame - 04
Speech & voice AI
Transcription, timestamped segmentation, speaker diarisation and audio event labels.
TranscriptsDiarisation
Explore Speech & voice AI
Mvaak transcription tool - 05
NLP & LLM data
Text classification, entities, intent, sentiment and span relations, with agreement checks.
EntitiesIntent
Explore NLP & LLM dataIllustrationBook a table for twoqty at 7 pmtime near Indiranagarloc.
intent: book_tablesentiment: neutral
- 06
Annotation & quality assurance
Labelling to a written guideline, multi-stage review and acceptance against agreed criteria.
GuidelinesReview
Explore Annotation & quality assurance
Real capture · unannotated
Collection regions
Based in India. Scoping projects across regions.
Field operations run from India. Collection elsewhere is scoped project by project — we confirm coverage before anything is promised.
Operations base
- India
Client engagements delivered
- United Arab Emirates
- Singapore
Target region — confirmed per project
- Philippines
- United States
- United Kingdom
- China
- Latin America (region)
- Other parts of Asia (region)
Technical capabilities
From capture to a training-ready dataset.
What is captured, labelled, checked and delivered — agreed per project in a written capture brief.
01
Capture
- First-person (egocentric) video
- Paired third-person views
- Calibrated RGB or stereo
- IMU clock-aligned to frames
- Multilingual audio & narration
02
Annotation
- Temporal segmentation to sub-action
- Outcome markers
- Hand keypoints & object tracks
- Contact events
- Task, object & environment metadata
03
Quality assurance
- File integrity
- Timestamp consistency
- Metadata & annotation completeness
- Task consistency
- Human review against your acceptance criteria
04
Delivery
- LeRobot
- RLDS
- MCAP
- Raw + manifest
- Bespoke schema
- Consent & provenance manifest
Institutional support
We are backed by
Awards & Recognitions
Ceremony — IndiaAI Application Development Pillar 
IndiaAI Innovation Challenge Award
Awarded to Narottam Innovations Pvt Ltd, presented with the mvaak AI name. More about Mvaak
Mvaak Partnerships
Build With Mvaak.
Partner with us to execute real-world data projects through structured workflows, training, and equipment support where applicable.
Who can apply
- Early-stage startups
- Technical teams
- Data collection operators
- Founding teams
What Mvaak may provide
- Equipment accessCapture devices may be provided where agreed and available.
- Practical trainingLearn the agreed workflow before work begins.
- Clear specificationsDocumented tasks, deliverables and quality criteria.
- Future opportunitiesPartners who deliver well may be considered for more work.
Applying is free. Every collaboration is agreed in writing first.
How it works
- 1Apply
- 2Evaluation
- 3Discuss the terms
- 4Training & setup
- 5Execute
- 6Review & grow
Real-world data for AI: egocentric (first-person) video of human tasks with synchronised sensor data, plus audio, video and text data with annotation and quality assurance.
Video recorded from a first-person perspective — typically a head- or chest-mounted camera — showing a task from the viewpoint of the person performing it.
We turn a task your robot needs to learn into a capture brief, record real people performing it in real environments, segment and annotate the episodes, and deliver them in formats such as LeRobot, RLDS or MCAP.
Household, commercial and agricultural settings, including retail, food service, offices and warehouse work. Industrial and other sites are confirmed during project scoping.
Yes. Standard deliverables include calibrated RGB or stereo video at the agreed resolution and frame rate, with IMU aligned to the video. The rig is chosen per task.
Yes. Tasks, environments, devices, annotation and delivery are defined in a written capture brief agreed with you before recording begins.
Yes. Annotation is done to a written guideline, and every delivery goes through integrity checks and human review against acceptance criteria agreed with you.
Mvaak runs a trained collector network across Indian states. Volume, timelines and staffing for a specific project are agreed during scoping — often starting with a small pilot.
Operations are based in India. Collection in other countries and regions is scoped and confirmed project by project.
Use the contact form and choose what you are looking for. Whether a sample is available, and what it contains, depends on the data and your requirements — we confirm this when you get in touch.
Start a project
Tell us what your robot needs to learn.
AI labs, robotics companies and research teams: describe the task and where it happens — we’ll help shape the dataset to train on.