Skip to content
AMAKH

Real-World Data for Physical AI.

From egocentric and industrial capture to robotics-ready annotation and evaluation, AMAKH helps AI teams turn real-world activity into useful training data.

Discuss a Data Project Explore Our Capabilities

Every frame carries layers of data.

We build each dataset in stages, from raw capture to accepted delivery. Every layer is captured, labelled and checked to your specification.

  1. BoxesPoseLabelsMetadata Accepted · your criteria
    Stage 01/05 · Capture
  2. BoxesPoseLabelsMetadata Accepted · your criteria
    Stage 02/05 · Sense
  3. BoxesPoseLabelsMetadata Accepted · your criteria
    Stage 03/05 · Track
  4. BoxesPoseLabelsMetadata Accepted · your criteria
    Stage 04/05 · Label
  5. BoxesPoseLabelsMetadata Accepted · your criteria
    Stage 05/05 · Verify

Illustrative frame

Discuss a data project

From egocentric and industrial capture to robotics-ready annotation, AMAKH turns real-world activity into data your models can learn from: captured with consent, checked against your criteria.

Grasp · fragile object

  • Physical AI & Robotics Data

    Real-world capture and data preparation for robotics and embodied AI.

  • Egocentric Video Collection

    First-person recordings of real tasks at home, at work and on the floor.

  • Warehouse & Industrial Data

    Forklift, material-movement and manufacturing workflows, captured with site consent.

  • Data Annotation & QA

    Task-specific labelling, review and consistency checks against your acceptance criteria.

Run as one program or on their own, with evaluation support alongside.

Explore capabilities

Real-world data for the places Physical AI has to work.

Capture and annotation programs scoped per site and per task. Outcomes depend on your models and how the data is used.

  • Robotics & embodied AI

    Task recordings and interaction data for manipulation, task learning and robotics evaluation.

    • Task recordings
    • Object interaction
    • Pose

  • Logistics & warehousing

    Forklift, picking and material-movement footage for warehouse automation.

    • Forklift workflows
    • Picking
    • Aisle movement

  • Manufacturing

    Assembly, inspection and line-side workflows recorded with site consent.

    • Assembly
    • Inspection
    • Line-side work

  • Household tasks

    First-person recordings of everyday tasks for human activity understanding.

    • Egocentric video
    • Everyday tasks
    • Hand-object

  • Industrial automation

    Real workflow data to study, train and evaluate automation in industrial settings.

    • Workcell footage
    • Process steps
    • Evaluation sets

Working somewhere else? Most capture plans start from the task, not the sector.

Discuss a data project

What we deliver

Five capabilities, used on their own or run together as one program from capture plan to delivered dataset.

Videos play on hover, or while in view on touch screens.

How we work

Physical AI learns from the physical world, and the physical world is hard to record well. AMAKH makes real-world data dependable: captured with consent, documented clearly and checked against the criteria you set.

  • Consent first

    No capture without the site’s permission and each participant’s consent.

  • Quality defined by you

    Acceptance criteria are agreed before work starts, and every delivery is checked against them.

  • Programs that adapt

    Pilots first, then repeatable programs scaled to what the environment, equipment and staffing allow.

  • Plain scope

    We say clearly what we can deliver and what we cannot, before the project begins.

From requirements to a delivered dataset

  1. Step 01

    Define requirements

    Agree what the model needs to learn: tasks, environments, devices, resolution, frame rate, duration and acceptance criteria.

  2. Step 02

    Plan capture

    Sites, participants, instructions, equipment and permissions are planned before anyone presses record.

  3. Step 03

    Collect data

    Field teams run the capture program, logging each session with the metadata your pipeline expects.

  4. Step 04

    Annotate & validate

    Trained operators label to your guidelines; samples are reviewed and corrected before anything ships.

  5. Step 05

    Deliver dataset

    Structured, documented data delivered in the format and cadence agreed for the project.

From the field

Homes, kitchens, warehouses and factory lines: the places Physical AI has to work, and the capture setup we typically plan for each.

  1. Folding and tidying

    View
    Fixed + first-person
    Task
    Laundry folding
    Layers
    RGB · Hand pose · Actions

  2. Food preparation

    View
    First-person
    Task
    Cutting and prep
    Layers
    RGB · Hand pose · Objects

  3. Picking and handling

    View
    Fixed + body-worn
    Task
    Carry and check
    Layers
    RGB · Body pose · Actions

  4. Line-side assembly

    View
    Fixed, multi-view
    Task
    Engine assembly
    Layers
    RGB · Depth · Actions

Typical setups, planned per project

Tell us what your model needs to learn.

Share the tasks, environments and volumes you have in mind. We will come back with a capture approach and the questions we need answered.

Start a project brief

Discuss a data project

The more you can share now, the more specific our first reply can be.

Useful to include: the tasks to record, environments (home, workplace, warehouse, factory), rough volume and timeline, and any device or annotation requirements.

All fields are required unless marked optional.