This paper introduces "snap-together motion", which is designed for interactive environments where the extensive graph search required to path-plan in their previous motion-graphs paper would be too time-consuming. In particular, they trade off absolute adherence to the original motion captures with a more compact graph -- one that approximates the "move trees" used in computer games.

Hub-based graphs (like move trees) are nice because they make finding transitions much easier. They have a high branching factor, but low depth -- which makes for efficient searches. And because each hub has lots of edges transitioning out from it, they provide for a neutral pose with numerous options for subsequent movement.

To build this graph, they construct hub nodes through automated transition point finding and with user-assistance. Once complication is that the merged nodes may not be exactly the same (though they will be similar), so some amount of blending is required to maintain both C0 and C1 continuity. In the process of blending, constraints may be violated, and they take care of the most common type: foot plants.