Skip to content

Human-in-the-loop data collection · Physical AI & robotics

Leverage Physical Superintelligence, with real-world data.

Humaid collects real-world, egocentric demonstration data for robotics — across manufacturing, warehouse, hospitality, and food service environments. Calibrated multi-sensor capture, trained operators, annotation, QC, and pipeline delivery.

Explore the data (opens in a new tab)
In the field · 20 rigs recording, 50 more deploying
  • Commercial kitchen line where chefs prepare and plate food at cooking stations.
    Commercial kitchen
  • Appliance assembly line with operators fitting panels along a conveyor.
    Appliance assembly
  • Nursery worker potting shrubs under a greenhouse canopy.
    Greenhouse
  • Electronics assembly workstations with operators working under articulated bench lamps.
    Electronics assembly
  • Construction site skyline with several tower cranes over a concrete structure.
    Construction site
  • Operator working at a machine with a touch-screen control panel on a production floor.
    Machine tending

Exclusive pilot program

Humaid launches its groundbreaking exclusive pilot program with one of the largest companies in Central Asia, AKFA Group — live capture across factories, logistics, construction, hospitality and retail sites.

0+
Employees in the group
0+
Companies in the group
0+
Countries of operation

What Humaid offers

Seven synchronized streams. One file.

Every session is multimodal by default. All signals are timestamped on one clock and delivered in a single MCAP container — ready for policy training without a stitching step.

Inspect a session in Data Explorer (opens in a new tab)
01 / 07 · Camera setup

How it works

From the factory floor to your training loop.

One vertically integrated pipeline. We do not crowdsource annotations or repurpose internet video — every dataset is collected on-site with calibrated equipment, trained operators and human review at every gate.

See the platform
  1. 01Protocol

    Protocol & rig design

    We define the task list, environments, sensor configuration, annotation schema and target volume with your team — then calibrate a rig for it. Onboarding a new dataset is configuration, not a new codebase.

    • Task protocol
    • Sensor config
    • Annotation schema
    Two engineers testing a robot arm together in a bright lab.
  2. 02Capture

    On-site capture in real facilities

    Trained, consenting operators wear calibrated multi-sensor rigs during their actual shifts — factories, kitchens, warehouses, hotels. Twenty rigs record today and fifty more are being deployed; sessions are remotely monitored, health-checked and auto-cut hourly.

    • 20 rigs live · 50 more deploying
    • Remote monitoring
    • Hourly sessions
  3. 03Segmentation

    AI proposes, humans decide

    A vision model proposes work segments; annotators review every proposal on a millisecond-precision timeline. Nothing ships without a human gate, and random clips are re-audited for quality.

    • AI chunking
    • Human review
    • QC red flags
  4. 04Enrichment

    Depth, hands, body, actions

    Neural depth, 21-keypoint 3D hands with wrist-camera calibration, full-body mesh, and temporal action labels — every stream aligned on one clock, with provenance flags so you can filter what you trust.

    • Hand pose
    • 3D body
    • Action labels
  5. 05Delivery

    Validated, blurred, delivered

    Automated validators reconcile every chunk, faces are blurred, manifests are checksummed. Datasets land in your bucket or on Hugging Face as MCAP — or converted to HDF5, RLDS and LeRobot.

    • MCAP
    • HDF5 · RLDS · LeRobot
    • Face blur

Data source verticals

Real work. Eight verticals.

Access to large industrial environments where we run live pilots and capture rare edge cases at scale — people doing their actual jobs, not actors in a lab.

ManufacturingWarehouse
Electronics assembly
Commercial kitchen
Component line
Metal profiles
Plant floor
Appliance assembly
Racking & storage
Assembly line
Machine tending
Greenhouse
Construction site
Production hall
Assembly hall
Appliance hall
  1. 01 / 08

    Manufacturing

    Assembly, machine tending and inspection on live production lines for cement, steel, glass and other building materials.

  2. 02 / 08

    Supply Chain & Logistics

    Picking, sorting, packing and pallet handling in distribution centres and cross-border logistics hubs.

  3. 03 / 08

    Construction

    Material handling, tool use and assembly work on residential, commercial and industrial sites, under real safety constraints.

  4. 04 / 08

    Mining

    Equipment operation, maintenance and material handling at extraction sites — heavy, safety-critical work far from any lab.

  5. 05 / 08

    Hospitality & Tourism

    Housekeeping, room service and front-of-house routines at large hotel and resort venues.

  6. 06 / 08

    Food & Beverage

    Commercial kitchens and processing plants — prep, plating, packaging and cleaning at institutional scale.

  7. 07 / 08

    Retail

    Shelf stocking, checkout and back-of-store handling across multi-format stores, with real customer flow.

  8. 08 / 08

    Healthcare & Wellness

    Pharmaceutical production and clinical support routines — sterile handling, packaging and logistics in regulated environments.

Why us

Diverse, scalable, high-fidelity.

We deliver diverse, hard-to-access real-world datasets through exclusive pipelines and partnerships, tailored to your exact use case — so you get model-ready data that fits your roadmap and scales to production.

Dataset formats
8

industry verticals

Diverse

Factories, warehouses, kitchens, hotels, construction and mining sites — real shifts with real variability, lighting changes and the long tail of edge cases a lab never produces.

20

rigs in the field · 3 cameras each

+50 deploying soon

Scalable

A fleet of remotely monitored capture rigs — 20 recording today, 50 more being deployed — with hourly auto-cut sessions and a durable, config-driven pipeline that turns terabytes per day into reviewed, labeled clips.

23

channels on one clock

High-fidelity

Stereo RGB-D at 1920×1200, three IMUs at ~200 Hz, hand and body pose, action labels — every signal in one MCAP with factory calibration and 0.00 ms depth-to-frame pairing.

Open data

See the data before you buy it.

Inspect real sessions in the Data Explorer and download an open dataset — the same MCAP files we deliver to customers.

Insights

Field notes on robotics data.

All articles