This paper addresses the problem of selecting and adapting a motion in order to make a virtual human interact with a user in real-time. We describe a method that is able to retrieve the most appropriate motion in interactive environments where both the constraints and the character's morphology are not known in advance. The method described here is based on a morphology-independent representation of motion to perform fast motion retargetting and retrieval. Because the size of the database is not infinite, even the best candidate motion may not satisfy precisely all the imposed constraints. As a second step of the algorithm, the selected motion is locally adapted in order to accurately satisfy these constraints. This adaptation allows dealing with a large set of possible movements even if no candidate in the database exactly fits the constraints.