What am I looking at?

This is the Motion Intelligence Codebook: a vocabulary of 10,000 different ways that things move.

Each small line represents a Motion Token, learned from real movement over time. Rather than describing what something looks like, each token captures a short pattern of motion: its direction, speed, acceleration and how those are changing.

Why are they arranged like this?

All Motion Tokens are overlaid on top of each other, with their directions initially aligned. Similar movements naturally sit near each other — green indicates speeds that are obviously representing some form of motorised movement, white represents speeds closer to human movement. Yellow are the in-betweens: fast joggers, cyclists, personal mobility devices — and indeed fast objects transitioning to faster or slower movement.

As the model learned its vocabulary, patterns emerge. Steady movement, turns, hesitation, acceleration and more unusual changes in motion occupy different neighbourhoods within the codebook.

The 3D view lets us explore that learned structure. It is a map of kinematic similarity, not a map of the physical world. Use your keyboard arrow keys to cycle through the most prevalent or the most rarely-used tokens. Click and drag to spin, and shift-click plus drag to move the Codebook.

Why Motion Tokens?

Language models turn words and text into tokens so they can learn patterns between them.

We apply the same principle to movement. By compressing space and time into Motion Tokens, a long stream of movement becomes a compact sequence the model can learn from, compare and predict.

Individual Motion Tokens describe movement. Sequences of Motion Tokens begin to reveal behaviour and intent. Think of these 10,000 Motion Tokens as the alphabet of motion. The intelligence comes from learning how that alphabet is used.