Skip to content
Menu

managed data collection

Real-world data for Physical AI.

Human demonstrations and real-world sensor data for teams training robots and Physical AI models. You define the tasks, environments and data requirements. We coordinate access, permissions, capture and quality review, so you get the data without running the fieldwork.

For teams worldwide · collection scoped by project

example brief

illustrative
task
Pump maintenance, from strip-down to reassembly
environment
Maintenance workshop
people
Qualified technicians
capture
First-person videoin scope
annotation
Task stepsoptional
calibration
Small batch, reviewed with youin scope

Illustrative only. Real briefs are agreed with you and confirmed as feasible before anything is recorded.

01 / why us

The hard part of real-world data, handled

Training the model is your work. Finding the right sites and people, and running capture well, is ours.

  • real

    Real work, real workplaces

    Tasks done by the people who normally do them, with the tools and conditions of a real site.

  • access

    No fieldwork for your team

    We find suitable sites and participants, arrange the visits and run the capture.

  • variety

    Variety by design

    Different people, sites, tools and conditions, planned into the collection rather than left to chance.

02 / process

How a collection works

You deal with one accountable team from brief to delivery, not separate sites, crews and contractors.

  1. Specify

    Tasks, environments, signals and acceptance criteria, defined by you and refined with us.

  2. Scope

    We confirm feasibility, access and permissions, then send a written proposal.

  3. Calibrate

    For larger projects, we propose a small calibration batch, reviewed with you before volume collection.

    Larger projects
  4. Collect and review

    Sessions follow a versioned protocol. Each one is reviewed, and rejection reasons are recorded.

  5. Deliver

    In the agreed format, with the specified metadata and documentation.

Share a data requirement

03 / signals

What the data contains, and what it doesn’t

We are plain about what human capture can and cannot provide. Apart from session metadata, every signal is agreed in your brief and confirmed as feasible during scoping.

  • Recorded for every session
  • Scoped per project
  • Not captured
  • Session metadataEvery sessionSession ID, equipment, protocol version and timestamps
  • First-person RGB videoScopedUsually the core signal, from a head- or chest-mounted camera
  • Task and step segmentsScopedDefined against your task taxonomy
  • Additional camera viewsScopedFixed cameras or a second wearable
  • AudioScopedAmbient sound or spoken narration of the task
  • DepthScopedFrom a depth-capable camera
  • IMU / head motionScopedFrom the wearable camera’s motion sensors
  • AnnotationScopedHands, objects or steps, defined with you
  • Robot joint states and actionsNot capturedNot present in human demonstrations
  • Force and torqueNot capturedNot measured by wearable capture

Packaging data in a robot-learning format such as LeRobot or RLDS organises what was recorded. It does not add signals that were never captured.

04 / environments

Wherever the task happens

If the work can be recorded safely and with the right permissions, we can scope a collection around it. Some examples:

  • env.a

    Building services and maintenance

    Electrical, plumbing, HVAC, servicing and repair

  • env.b

    Workshops and fabrication

    Machining, welding, joinery and finishing

  • env.c

    Assembly and manufacturing

    Assembly, kitting, inspection and line work

  • env.d

    Logistics and warehousing

    Picking, packing, handling and stock work

  • env.e

    Hospitality and kitchens

    Food preparation, cooking, service and cleaning down

  • env.f

    Facilities and cleaning

    Housekeeping, cleaning and site services

  • env.g

    Agriculture and food processing

    Harvesting, sorting, trimming and packing

  • env.+

    Your environment

    Describe the task and the setting. We will assess access, safety and permissions.

Examples of work we can scope, not existing datasets or site agreements.

05 / rights and quality

Recorded for every session, not assumed

Rights and quality evidence travel with the data, so your legal and data teams can check what they are receiving.

  • permissions

    Permissions first

    Participant, site and client permissions are in place before anything is recorded. People can decline.

  • licence

    Scope in writing

    Training use, territory and exclusivity are defined for each project.

  • provenance

    Traceable sessions

    Session IDs, protocol versions and checksums are delivered with the data.

  • acceptance

    Measured, not blended

    Captured, submitted and accepted units are reported separately.

for employers, teams and professionals

Run a team or a site? Become a collection partner.

Paid, scheduled recording projects in your own workplace. Permissions and compensation are agreed in writing before anything is recorded.

Have a data requirement?

Tell us what you’re training and what the data needs to contain.

Share a data requirement

or email contact@motion2machine.com

A high-level outline is enough to start.