### Summary
This paper introduces motion graphs as a method for connecting motion clips to fit user-specified constraints.
### Problem
Without very careful planning, it can be difficult to determine motions which can be blended well.
### Methods Used
To create the graph:

* A distance function is evaluated for every pair of frames in the motion database
* Local minima in this distance function are extracted and saved as good transition points
* Linear blending for the root and spherical linear interpolation for joint angles are used to create transitions between motions at these minima
* The graph is examined and pruned to remove paths that can't be reached or escaped

To use the graph:

* A user-specified error function and halting condition is given, and a graph walk found which minimizes the error while satisfying the halting condition
  * For the specific case of path synthesis, the error function is the squared distance between the taken and desired paths, and the halting condition is met when the character reaches his destination

### Key Ideas
Numerically close frames are good transition candidates
Graphs can be used to find good transitions between motions and to generate sequences of motions subject to constraints
Only good transitions should be generated automatically

### Questions
Toward the end of this paper, one of the applications mentioned was interactive control, for which graph walks would be precomputed and then replayed in response to user input.  How is this significantly better or less work than the prior technique of motion trees?

Also, is there any accounting for the possibility that a dead-end graph node could be a desirable pose (i.e., the user wants to finish his action in a frame with no transitions)?