This paper introduces the concept of a motion graph, which allows for creation of long sequences of motions from spliced clips from a motion database. This work extends their previous work with registration curves. The graph itself has clips represented by directed edges and transition points represented by vertices.
Transitions are precomputed for each pair of clips by looking for points in time where the poses are similar (given a user-specified threshold). Once found, each transition is then added to the graph as a new edge connecting previous clips. After the graph is built, some problems may remain: the graph may have motions that are logically discontinuous, and may also have dead end nodes. Both are solved by pruning all but the strongly connected components.
Now, as long as two poses are connected by a series of edges, there's guaranteed to be a directed path between them. The path itself can be found by a regular search, though they use a branch and bound strategy to optimize this search. Walking along this path produces the motion, with transition clips given as alpha-blended compositions of the source and target motions.