Motion Atlas — The Global Dataset for Physical AI. Motion Atlas builds structured, synchronized multi-camera human-motion datasets and motion intelligence for robotics and Physical AI.

The data problem

Intelligence is moving into the physical world.

Artificial intelligence has learned extensively from the digital world. Physical AI introduces a different challenge. A machine operating in the real world must understand more than what an object is.

The physical world must be experienced, not merely described.

• how objects are approached
• how they are grasped
• how they are manipulated
• how tools are used
• how two hands coordinate
• how environments influence movement
• how actions unfold through time
• how individual actions combine into meaningful tasks

Embodied knowledge

Human activity contains knowledge.

Everyday human activity contains an enormous body of physical knowledge. That knowledge appears in ordinary actions people have learned through experience. Much of it was never formally written down. It exists in how people act.

It exists in motion.

• folding a towel
• organizing a workspace
• preparing food
• using a tool
• handling unfamiliar objects
• cleaning
• assembling
• sorting, carrying, completing the work

Motion Atlas captures physical experience and transforms it into structured data designed to become useful for machines learning about the physical world.

How Motion Atlas works

From human activity to motion intelligence.

Motion Atlas captures real-world human activity as Episodes. An Episode is a coherent representation of a physical activity connected across viewpoints, time, media, task structure, objects, interactions, observable execution, quality, and provenance. Capture, synchronize, structure, understand, validate, and deliver — this is the public capability, not a list of internal applications.

One activity. One Episode. Multiple perspectives. One shared timeline.

1. 01

   Capture

   Real-world human activity, as observed

2. 02

   Synchronize

   One Master Time across views and annotations

3. 03

   Structure

   Episode, task hierarchy, actions, objects, contacts

4. 04

   Understand

   Observable execution context on the same clock

5. 05

   Validate

   Quality, provenance, protected originals

6. 06

   Deliver

   Motion intelligence for robotics and Physical AI

The Episode

More than video.

Video records what a camera saw. Physical intelligence requires context. Motion Atlas connects synchronized observations with task structure, time, objects, interactions, observable execution, quality, and provenance. Depending on the collection, an Episode may also incorporate richer motion representations and additional sensor-derived information. Not every Episode contains every optional layer.

The result is not simply a collection of recordings. It is a structured representation of human activity designed to become increasingly useful for Physical AI.

Task hierarchy

From movements to meaningful tasks.

A hand reaches. An object is grasped. A tool is positioned. An adjustment is made. Another action follows. A mistake may be corrected. Eventually, an objective is accomplished. Isolated movement clips cannot explain that. L1 atomic action. L2 skill. L3 activity. Semantic task boundaries matter more than arbitrary clip length.

Intelligent machines ultimately need to learn not only how people move, but how movement becomes useful action.

Diversity

The physical world has no single way of moving.

The same objective can be accomplished differently depending on the person, the environment, the objects, the tools, the workspace, and the technique. Homes differ. Workplaces differ. Objects differ. Tools differ. Techniques differ. People differ. A machine intended to operate beyond a controlled environment must eventually encounter this variation. Motion Atlas is being designed to capture human activity across different people, environments, objects, tasks and ways of accomplishing them — a collection objective, not a claim that the current catalog already represents the world.

The world is diverse. The data that teaches machines to operate within it should reflect that reality.

Quality

Scale matters. Trust matters more.

Large amounts of data can be valuable. Volume alone does not make physical data useful. Data becomes harder to trust when its origin is uncertain, its timing is unreliable, its quality is unknown, its transformations cannot be traced, or its relationship to the underlying activity has been lost. Motion Atlas is being designed around synchronization, structured metadata, quality assurance, versioning, traceability, provenance, and preservation of original media.

Quality should be measurable. Provenance should be traceable. Originals should remain protected.

Larger purpose

Building useful intelligence for the physical world.

The practical value of Physical AI will increasingly depend on what that intelligence enables machines to do in the physical world — in manufacturing, logistics, healthcare and care environments, agriculture, construction, maintenance, homes, and other forms of useful physical work. Motion Atlas does not claim to transform those industries. It exists to contribute one part of this larger undertaking: transforming knowledge embedded in human activity into motion intelligence that machines can learn from.

Mission

To map human motion at meaningful scale and transform it into motion intelligence that advances robotics and Physical AI for the benefit of people and society.

From human experience, to machine understanding, to useful action in the physical world.

Human Motion. Machine Intelligence.

Motion Atlas Platform

One motion intelligence platform. Specialized workspaces.

Nexus is this site — the public, research, and customer gateway. Dataset Explorer is where you inspect structured sample Episodes. Studio is a separate application for Episode operations.

FAQ

Straight answers before a license call.