Summary This paper describes a method for selecting a collection of motion capture clips to synthesize a new sequence of motion that satisfies a variety of hard and soft constraints. This method is used to solve the problem of generating smooth and natural-looking strings of motion that do not already exist in a motion capture clip collection. The algorithm proposed in this paper represents the problem of stringing together separate clips of motion as a graph search and uses and randomized search method to quickly generate multiple motions that all satisfy a given set of constraints. Key Ideas 1. The collection of motion capture data is represented as a directed graph where each node is an individual sequence of motion and each edge between two nodes indicates a pair of frames where a cut can be made from one motion to the other. Each edge is also assigned an edge weight whose cost is determined by a similarity metric representing the cost of discontinuity of connecting the incident frames. Cuts between motions that will introduce discontinuities carry a high cost. 2. Finding paths in the motion graph that satisfy the hard constraints and optimize soft constraints involves a graph search. To gather a set of results in real-time, the algorithm uses a randomized search working on multiple levels to first converge on a minimized path and then smooth out discontinuities. Questions/Comments The authors admit that the quality of the synthesized motions relies strongly on the integrity of the original motion dataset. I wonder if there is a way of pre-evaluating the strength of the motion dataset prior to using it for synthesizing new motions based on the desired constraints.