$0B
Humanoid robot market by 2035
Goldman Sachs
Nxted Capture
Robots can’t learn from the internet.
They learn from people.
LLMs had the whole internet to read. Robots have no internet of actions - the two largest open robot datasets combined add up to roughly 5,000 hours. Physical AI is bottlenecked on one thing: first-person video of real humans doing real, skilled work.
Nxted Capture records that data from India’s skilled workforce - at a fraction of Western cost, with the diversity the scaling laws reward.
Egocentric · first-person
Real hands. Real tools. Real workshops.
What is egocentric data for robotics?
Egocentric data for robotics is first-person video recorded from the worker’s own point of view as they perform a task - what a robot’s head- or wrist-mounted camera would see - plus depth, hand pose and 6-DoF motion. It teaches manipulation policies how a skilled human actually moves, which third-person or web video cannot.
Why capture this data in India?
India has one of the world’s largest skilled workforces across the trades and professions the scaling laws reward. nxted captures only verified, consented contributors in real workplaces, with fair pay and documented provenance. Skill depth and provenance - not cost - are the point, though Indian capture is also materially lower-cost.
The opportunity
The biggest bottleneck in robotics isn’t compute or models. It’s data of humans doing things.
$0M+
Spent yearly by robotics firms buying real-world data
MIT Technology Review
~5,000 hrs
The entire open-source robot dataset supply, combined
Scale AI
0%
Policy gain from human egocentric data vs robot data alone
Meta · EgoMimic
No internet of actions
Language models trained on trillions of web tokens. There is no equivalent corpus of physical manipulation - so robots have to be shown, demonstration by demonstration.
Diversity beats volume
Peer-reviewed scaling laws (ICLR 2025) show robot-policy generalisation follows a power law in the number of environments and objects - exactly what a diverse, India-wide workforce provides.
The money is already moving
Over $6B was invested in humanoid robotics in 2025. Robotics firms already spend $100M+ a year buying real-world data. The supply hasn’t caught up. That’s the gap.
How it works
From a skilled worker in a workshop to a training-ready dataset.
01
Source verified skilled workers
We recruit and verify tradespeople across India - tailors, machinists, carpenters, chefs, medical staff - matched to the skill and level you need.
02
Fit the capture rig
Workers wear egocentric (first-person) glasses built on research-standard devices, plus depth and hand-pose sensors for robotics-ready tiers.
03
Record real tasks, real environments
We film the actual job in the actual workshop - not a sterile lab. Diversity of scene and object is what the scaling laws reward.
04
Annotate & segment
Action segmentation, first-person narration, hand-pose tracks, 6DoF trajectory, and skill-level metadata - the Ego4D / Ego-Exo4D methodology.
05
Quality & consent review
Every batch is reviewed by our expert team. Signed releases, PII blurring, and a GDPR-compliant DPA cover every frame before delivery.
06
Deliver in your format
LeRobot, RLDS, HDF5, or MP4 + sidecars - dropped straight into your imitation-learning or VLA training stack.
The capture rig
Built on the devices the research field already trusts.
We don’t ship fictional proprietary hardware. Our rigs are built on Meta Project Aria, Intel RealSense, Stereolabs ZED, and the Universal Manipulation Interface - the same stack behind Ego-Exo4D and EgoMimic. Three tiers, matched to how training-ready you need the data to be.
First-person RGB
up to 1408² / 30fps
Depth + point cloud
RealSense / ZED
Hand pose
per-joint tracking
6DoF trajectory
SLAM-derived
Eye gaze
attention signal
Action labels
segmented + narrated
Entry
High-volume, low-cost
Sensors
- ›Head-mounted 4K RGB camera
- ›Built-in IMU (accel + gyro)
Outputs
- ›1080p / 30fps video
- ›Motion telemetry
- ›Action timestamps + narration
Formats
- ›MP4 + JSON/CSV sidecars
- ›Convertible to LeRobot
Pre-training video corpora; scaling across many workers fast
Built on UMI-style GoPro capture
Standard
Most popularResearch-grade egocentric
Sensors
- ›Aria-class glasses: RGB + SLAM cameras
- ›Eye tracking
- ›Dual IMU + spatial audio
- ›Optional fixed exo camera
Outputs
- ›RGB up to 1408² / 30fps
- ›6DoF SLAM trajectory
- ›Eye gaze + hand tracking
Formats
- ›LeRobot (Parquet + MP4 + JSON)
- ›MP4 + sidecars
The human→humanoid recipe - same glasses on worker and robot
Built on Meta Project Aria
Robotics-ready
Direct policy training
Sensors
- ›Everything in Standard
- ›RealSense / ZED depth
- ›Hand-pose mocap gloves or UMI gripper
- ›Multi-view exocentric
Outputs
- ›RGB-D + point clouds
- ›Per-joint hand pose
- ›Gripper width / contact
- ›6DoF trajectory + gaze
Formats
- ›LeRobot
- ›RLDS / TFDS
- ›HDF5 (ALOHA / robomimic)
Imitation learning & VLA training; Open X-Embodiment-ready
Built on Aria + RealSense + DexCap / UMI
The 5 levels
From basic packing to specialist craft.
Level 01
Foundation
from $35/hrSorting, stacking, packing, basic assembly, labeling
Example task
Pack 50 items per minute across 5 product types
Level 02
Skilled Trades
from $55/hrTailoring, carpentry, cooking, cleaning, gardening
Example task
Complete garment assembly with hand and machine techniques
Level 03
Technical / Industrial
from $80/hrCNC operation, welding, electrical work, plumbing, electronics assembly
Example task
Set up and operate a CNC lathe for precision part manufacture
Level 04
Professional / Medical
from $120/hrSurgical assistance, patient care, pharmacy, dental, lab work
Example task
Prep and assist in laparoscopic procedure, instrument handling
Level 05
Rare / Specialist
Custom quoteHeritage crafts, traditional medicine, precision jewellery, instrument making
Example task
Traditional Kanjivaram silk weaving - 40-step process
Formats
We deliver in formats your robotics pipeline already uses.
MP4 4KRLDSHDF5LeRobotCustom
Every dataset includes
- ✓First-person (egocentric) perspective
- ✓Action timestamps and task segmentation
- ✓Worker skill level and experience metadata
- ✓Environment classification
- ✓Ethics consent documentation (GDPR-compliant)
Optional enrichment
+ Depth maps+ Pose skeletons+ Semantic segmentation+ Optical flow+ Hand tracking+ Action labels
Sample dataset
What actually lands in your bucket.
Every dataset ships in a consistent, robotics-ready structure - raw and processed video, multi-format episodes, full metadata, and the compliance pack. Below is the shape of a typical industrial delivery.
Example structureIllustrative
nxted_industrial_india_01/ dataset_card.md # scope, splits, limitations consent_manifest.csv # per-contributor consent + pay task_cards/ # what each episode demonstrates raw_video/ # original egocentric capture processed_video/ # redacted, stabilised, trimmed exo_video/ # third-person reference angle metadata/ # calibration, 6DoF poses, timestamps annotations/ # action labels, segments, success quality_report.pdf # inter-annotator agreement + QA lerobot/ # LeRobot-format episodes rlds/ # RLDS / TFDS shards hdf5/ # HDF5 trajectories
Every dataset includes
- Raw and processed egocentric video, redacted to spec (faces, plates, screens, PII).
- Third-person reference video where a second angle helps the policy.
- Per-episode task cards plus action labels, segments and success / failure flags.
- LeRobot, RLDS and HDF5 exports of every episode - drop straight into your pipeline.
- Full metadata: camera calibration, 6DoF trajectories, hand poses and timestamps.
- Dataset card and data-provenance log tracing every clip back to its source.
- Consent manifest and fair-payment record for every contributor.
- QA report with inter-annotator agreement and labelled edge cases.
The consent, provenance, redaction and QA artifacts are bundled as the Data Trust Pack - the documentation your data, legal and safety teams need to sign off a dataset for production.
Proof it works
The field has already shown human capture trains better robots.
We don’t ask you to take our word for it. Here is what the leading labs and companies have published - every claim links to its source.
Skill categories
What we can record.
Our flagship vertical is skilled industrial and technical work - electrical panel assembly, machine tending, CNC setup, electronics assembly, and inspection - the data the free open datasets don't cover and gig networks can't reach.
Electrical / Industrial assembly
Flagship verticalPanel wiring · Electronics assembly · Machine tending · Inspection
CNC / Machining
Skilled machinistsLathe ops · Mill setup · Quality inspection
Tailoring & Textile
Deep skilled poolHand stitching · Machine sewing · Pattern cutting
Construction / Warehouse
Heavy physical tasksBricklaying · Tile setting · Pick and pack
Carpentry & Furniture
Complex 3D assemblyJoinery · Lathe work · Lacquer finishing
Cooking & Food
Multi-cuisine poolKnife skills · Plating · Tandoor / griddle
Medical & Surgical
Credentialed cliniciansSuturing · Instrument handoff · Patient prep
Why India
The skill depth of centuries of craft - captured with consent.
Skill
Real tradespeople doing real, skilled work - electrical assembly, machine tending, CNC setup, tailoring, surgical-adjacent tasks. The breadth of credentialed trades the scaling laws reward, and that Western lab-bound datasets lack.
Trust
Workers paid above local market rate, fully consented, with signed releases and PII/face/plate/screen redaction. Every frame is covered by a DPDP and GDPR-aligned DPA. Consent-first capture is the product, not an afterthought.
Provenance
Every dataset ships with a Data Trust Pack: consent records, skill verification, reviewer credentials, a dataset card, and a QA report. You always know who produced the data and how.
Cost
Indian capture is also materially lower-cost, but cost is the footnote, not the pitch.
Further reading & sources
Scale AI - Expanding the Data Engine for Physical AIMeta - Ego-Exo4D: skilled human activity datasetGoogle DeepMind - Open X-Embodiment (1M+ trajectories)Data Scaling Laws in Imitation Learning (ICLR 2025)Goldman Sachs - humanoid robot market forecastMIT Technology Review - the humanoid data gig economyHugging Face - the LeRobot dataset formatMeta - Project Aria Gen 2 research glasses
Market forecasts are quoted as attributed estimates and diverge across analysts. Cost differentials are industry estimates presented as ranges. Research citations link to primary sources (company sites, arXiv, dataset homepages).
Tell us the skill. We’ll capture the data your robots need.
Custom datasets quoted within 24 hours, in any format your pipeline already uses.
FAQ
Egocentric data for robotics: FAQ
What is egocentric data and why do robots need it?
Egocentric data is first-person video recorded from the doer’s point of view - what a robot’s own camera would see - usually with depth, hand pose and 6-DoF motion. Robots need it because there is no web-scale corpus of physical actions; manipulation policies have to be shown how skilled humans move, demonstration by demonstration.
What hardware does nxted capture with?
nxted builds on the rigs the research field already uses: Meta Project Aria glasses, Intel RealSense and Stereolabs ZED depth cameras, and Universal Manipulation Interface grippers - the same class of stack behind datasets like Ego-Exo4D. Three tiers map to how training-ready you need the data, from RGB-only up to full depth and hand pose.
What skill categories can nxted record?
The flagship vertical is skilled industrial and technical work - electrical panel assembly, machine tending, CNC setup, electronics assembly and inspection. nxted also captures tailoring and textile, construction and warehouse, carpentry, cooking, and medical-adjacent tasks, using credentialed contributors for the specialist categories.
What formats and annotations are included?
Each dataset ships in LeRobot, RLDS and HDF5, plus raw and processed egocentric video, an optional third-person reference angle, full metadata (camera calibration, 6-DoF trajectories, hand pose, timestamps), action segmentation and success/failure labels, a dataset card, a consent manifest and a QA report.
Is nxted’s capture consented and ethically sourced?
Yes. Contributors are paid above the local market rate, give explicit, withdrawable consent, and every filming site signs a release. Faces, plates, screens and PII are redacted, no contributor is under 18, and each delivery includes a Data Trust Pack with a DPDP & GDPR-aligned DPA.
How much does an egocentric dataset cost?
Start with a Physical AI Test Kit from $2,500 - 5 to 10 usable hours of one skilled task with a consent pack, metadata, basic labels and a LeRobot/RLDS/HDF5 sample, delivered in 7-10 days. Full datasets are priced per usable hour by skill level and quoted within 24 hours.