Skip to content

THOR / Physical intelligence network

Machines, Builtfor Humanity.

THOR is an open source robotics lab building data, software and hardware for the next industrial revolution.

See how it works
  • Open market
  • Verified output
  • Global participation
  • Physical AI
  • Open market
  • Verified output
  • Global participation
  • Physical AI
  • An operator at a workstation driving a pair of robot arms over a bench of small objects.

    REC / 01

    ROBOT STATE + CONTROL

    An operator drives the robot. Cameras, joints and gripper signals are recorded together as clean action ground truth.

  • A handheld gripper rig held over a dish rack, recording the task from the wearer's point of view.

    REC / 02

    FIRST-PERSON DEMONSTRATION

    A person performs the task in the real world. A handheld gripper captures movement close to the robot's future point of view.

  • An annotator reviewing recorded robot footage frame by frame on a desktop monitor.

    REC / 03

    LABELS + QUALITY CONTROL

    People interpret the hard parts: action boundaries, contact, failure, intent and the cases where automated labels are uncertain.

0142
0142

Annotation example / 01

See a task becometraining data.

Play the episode, scrub through each subtask and switch the visual masks on or off. The timeline shows how THOR turns one continuous robotics run into measurable, verifiable units.

A gold-foiled satellite on a test rig, two robot arms aligned against it under studio lights.

THOR / ANNOTATION RECORD 0142

SUBTASK

✓ Initialise

0:00

An operator driving a pair of robot arms across a workbench while overhead cameras record the run.

Live annotation view / 01

Robotics annotation data types

Examples of the training signals customers can request through the network.

  • ACTION GROUND TRUTH

  • PROPRIOCEPTION + TACTILE

  • TIME SYNCHRONISATION

  • TEMPORAL SEGMENTATION

  • SEMANTIC CAPTIONS

  • QA + FAILURE FLAGS

  • 3D HAND POSE

  • HAND-OBJECT CONTACT

  • GAZE + ATTENTION

Helping intelligent machines learn faster, adapt better, and move with purpose.

Live annotation view / 01

One raw capture.Four layers of proof.

Step through the same task as THOR turns footage into measurable training signal. This interaction is the visual centre of the brand because the label, not the robot render, is the product.

THOR / TH-0187

F / 000214 · T 04.20S

Two robot arms over a kitchen island, one reaching into an open cutlery drawer while the other holds a spatula above a bowl.DRAWER · 0.97

ACTION / 03

reach → grasp → pull

  • Reach
  • Contact
  • Pull
  • Open
  • ORDER / LIVE SPEC
  • INPUT / VIDEO + STATE
  • OUTPUT / FRAME-ALIGNED LABELS
  • MODE / VERIFY

ACCEPTANCE RULE

4 / 4 GATES PASS

OUTPUT ACCEPTED

The problem / 01–03

The hard part startsbefore training.

A six-axis industrial robot arm drawn as a wireframe mesh.

01-RBT-001

Active unit

01-RBT-007

Motion system

  • 01 / 03

    The record wasnever made

    Language AI trained on an internet that already existed. Robot experience has to be recorded deliberately, one task and environment at a time.

  • 02 / 03

    Raw footage is nottraining data

    A useful hour needs action boundaries, object state, contact, timing and a shared definition of what counts as complete.

  • 03 / 03

    The market is closed

    Every new specification restarts sourcing, samples, terms and pilots. Small labs and specialist suppliers struggle to find each other.

The brand idea

Machines arenot the oppositeof people.

Physical intelligence only becomes useful when people record, interpret and verify the world it learns from. THOR makes that contribution visible.

Our problems are man-made, therefore they may be solved by man.
John F. Kennedy
A humanoid robot and an older man sitting together on a sofa in a domestic living room.

The response / THOR

Precise demand.A world ofcapable supply.

  1. 01 / Demand

    Post the specification

    Source data, required output, budget and acceptance rule.

  2. 02 / Production

    Compete on output

    Miners produce work with people, models or hybrid systems.

  3. 03 / Proof

    Verify the result

    Validators test submissions against the posted requirement.

  4. 04 / Delivery

    Deliver what passes

    Accepted data returns under the rights defined by the order.

Who does what

One market, four accountable roles.

01 / 04

Validated output

Only accepted data returns to the customer.

The expansion

Three stages.One open foundation.

THOR starts with physical annotation, then develops models and open-source hardware before testing the system with industry partners.

  • 01

    Stage 01

    Physical Annotation

    People transform existing video, sensor and robot data into verified training signals.

    Data standard

  • 02

    Stage 02

    Models & Hardware

    Train and test models on verified data, and develop open-source hardware for capture and deployment.

    Open systems

  • 03

    Stage 03

    Industry Pilots

    Work with industry partners to test models and hardware in operating environments.

    Deployment

A humanoid robot at an outdoor workbench showing a build to two children, a robot dog waiting on the pavement beside them.

THR / OPEN_WORLD_015

The human part

More capable machines begin with human work.

People demonstrate tasks, record what happened and check whether the result is useful. THOR connects that work to the robotics teams building the machines.

Physical AI should be built with the world, not merely delivered to it.

Build physical intelligence in the open.

Investor centre