This paper describes a method that can significantly increase the size of a collection of motion observations by cutting limbs from one motion sequence and attaching them to another. Not all such transplants are successful, because correlations across the body are a significant feature of human motion. The method uses randomized search based around a set of rules to generate transplants that are (a) likely to be successful and (b) likely to enrich the existing motion collection. The resulting frames are annotated by a classifier to tell whether they look like human motion or not. We evaluate the method by obtaining motion demands from an application, synthesizing motions to meet those demands, and then scoring the synthesized motions. Motions synthesized using transplants are generally somewhat better than those synthesized without using transplants, because transplanting generates many frames quite close to the original frames, so that it is easier for the motion synthesis process to find a good path in the motion graph. Furthermore, we show classifier errors tend to have relatively little impact in practice. Finally, we show that transplanted motion data can be used to synthesize motions of a group coordinated in space and time without producing motions that share frames.