TazQ
Academy Let’s connect

04 / Capability

Teach machines how the physical world works.

TazQ provides Physical AI data: first-person (egocentric) recordings of skilled people doing real tasks, human demonstrations for robot learning, and expert annotation of video, LiDAR and multi-sensor data. Teams building robots, wearable assistants and world models use it to learn how hands, objects and environments interact.

Plan a Physical AI project Explore the 3D scene

Updated Robotics · LiDAR · World models · Egocentric video

Definition

What is Physical AI data?

Physical AI data is the training and evaluation data for AI systems that perceive and act in the real world, such as robots, autonomous machines and wearable assistants. It pairs recordings of how people perform tasks with precise labels for objects, hands, actions, depth and motion, so a model can connect what it sees with what it should do next.

Why it matters

Why the physical world needs human data.

Models learn to act from examples. For the physical world, the right examples rarely exist online.

01

The internet rarely shows how work is done.

Web video seldom captures the hand-level detail, the viewpoint and the exact sequence a robot needs to repeat a task reliably.

02

Sensors need human context.

Point clouds and multi-camera streams become useful only when someone defines what each object is, where it is, and what is happening to it.

03

Edge cases live in real places.

Clutter, glare, reflections and unusual objects appear in real homes, kitchens, workshops and warehouses, not in clean lab scenes.

Free: the Physical AI readiness checklist

Seven decisions for robotics and embodied AI data, from which tasks to capture to how every sensor lines up in time.

Interactive / Annotation layers

See a task through the expert’s eyes.

One reach-and-grasp task, captured as a 3D scene. Switch on the layers an annotation team adds, play the action, and drag to look around.

  1. Reach
  2. Grasp
  3. Lift
  4. Place
Illustrative scene. Each layer is a separate annotation pass: raw points, depth, 3D object cuboids, the hand’s trajectory and timestamped actions.

How TazQ does it

From a real task to model-ready data.

Every project follows the same four steps, shaped around the task your system must learn.

  1. 01

    Scope the task and its world.

    Define the activity, the environments it happens in, the variation the model must handle, and how success will be measured.

  2. 02

    Capture real people doing real work.

    Trained contributors record tasks from a first-person view or on multi-sensor rigs, following written protocols for setup, pacing and coverage.

  3. 03

    Annotate every layer the model needs.

    Objects, hands, depth, 3D cuboids and timestamped actions are labelled in separate passes, each against its own guidelines.

  4. 04

    Review, measure and deliver.

    Expert review and agreement checks run on every batch before delivery in the formats your training pipeline already uses.

Deliverables

What you receive.

Formats adapt to your pipeline. These are the data types teams ask for most.

Physical AI data types, what each includes, and typical formats
DataWhat it includesTypical formats
Egocentric videoFirst-person recordings of real tasks, with task descriptions, environment and device metadata.Video (e.g. MP4) with JSON metadata
Human demonstrationsComplete task executions for imitation learning, segmented into steps and synchronised with sensor logs.Video, step segments, sensor logs
Hand and body poseFrame-by-frame keypoints and tracks for hands and bodies, including occluded moments.Keypoint JSON (e.g. COCO-style)
Object and scene labels2D boxes, segmentation masks and 3D cuboids for objects, surfaces and free space.COCO, mask files, cuboid JSON
LiDAR and point clouds3D cuboids, semantic point labels and object tracks across frames.Point cloud files (e.g. PCD, LAS) with labels
Action and intent labelsTimestamped verbs, objects and goals, such as “grasp mug” or “open drawer”.Timeline JSON or CSV
Evaluation setsHeld-out scenarios and human judgments for testing robot policies and world models.Scenario specs with scored results

Want to see one?

Request an anonymised sample of any deliverable. A specialist will share a relevant example where available.

Use cases

Where teams use it.

01Household and service robotics

Manipulation in cluttered, real homes: grasping, opening, sorting and tidying, demonstrated by many people across many spaces.

02Warehouse and industrial automation

Picking, packing and inspection footage with object labels and the failure cases that busy, imperfect environments produce.

03World models and video prediction

Long, varied first-person sequences with action labels, so models learn what tends to happen next and why.

04Wearables and AR assistants

Egocentric video with hands, attention and intent, so an assistant understands what the wearer is doing and what they need.

Quality & experts

Quality starts with people.

Trained contributors

Contributors are prepared through TazQ Academy before they record or annotate, so protocols are followed the same way across people and places.

Explore TazQ Academy

Consent and privacy by design

Participants know how their recordings will be used, and faces, screens and personal details in the background are anonymised before delivery.

Layered review

Every batch passes automated checks and expert review against written guidelines. Agreement is measured, and what reviewers learn feeds back into the guidelines.

FAQ

Questions, answered.

01

What is egocentric data?

Egocentric data is video recorded from a first-person point of view, usually with a head- or chest-mounted camera, so it shows what a person sees while their hands work. It is the closest viewpoint to a robot’s or a wearable device’s own camera, which makes it especially useful for learning manipulation and everyday tasks.

02

How is Physical AI data different from regular computer vision data?

Computer vision data often labels single images. Physical AI data follows a task through time and space: it links objects, hands, depth, 3D position and actions across a sequence, often from several sensors at once, so a model can learn what to do, not only what it sees.

03

Which sensors and formats can you work with?

Projects can include RGB and depth cameras, multi-camera rigs, LiDAR and motion sensors. Data is delivered in the formats your pipeline uses, for example video with JSON metadata, COCO-style annotations or labelled point cloud files.

04

Can you capture data in specific environments?

Yes. Each project is scoped around the environments your system will operate in, such as homes, kitchens, workshops, retail spaces or industrial sites, and around the variation it must handle, from lighting to clutter to object types.

05

How do you protect the privacy of people in recordings?

Participants take part with informed consent, and recordings are reviewed so that faces, screens and personal details that are not part of the task are anonymised before delivery.

06

How do you measure annotation quality?

Each layer has written guidelines and reference examples. Batches are checked automatically and reviewed by experts, agreement between annotators is measured, and results are reported with every delivery.

07

How do we start a Physical AI project?

Share the task, the target environment, the sensors you use and the volume you expect. A small pilot dataset then validates the protocol, guidelines and format before the project scales. You can outline all of this in the project brief below.

Related capabilities

All capabilities
03 /↗

Multimodal & Speech AI

Language, voice, image, and video data with the context models need to connect what they see, hear, and understand.

  • ASR
  • TTS
  • VLM
02 /↗

Agentic AI

Human-led trajectories, realistic task environments, and deep evaluation for agents that need to act, adapt, and complete real work.

  • Trajectories
  • RL environments
  • Evaluation

Start your Physical AI project

Let’s build
what comes next.

A better model starts with a clearer question.
What do you want yours to understand?

For AI teams

Scope your project in about a minute. A specialist reads every brief and replies with next steps.

For human expertsExplore TazQ Academy
YOUR PROJECT, IN FOCUS
Who are you?
What are you building?

Choose all that apply.

What brings you here now?

Choose up to two.

Where are you today?

Stage

When would you like to start?

Scale (optional)

Tell us about the challenge

Languages or domains (optional)

Send your brief to a specialist

Your role (optional)

We’ll only use this to reply about your project. Please don’t include confidential details.