Abstract: When creating animated characters, we wish to develop models of motion that can generalize beyond a specific set of recorded motions. Significant progress has been made towards this end, including the development of controllers of growing sophistication for physically-simulated characters. In this talk, we take stock of some of the recent advances in the control of locomotion, with an eye towards identifying some of the common elements that might help define a future era of "downloadable skills." We argue in favor of learned memory-based models that allow for skills to be developed and integrated in an incremental fashion.