Animated characters are a crucial component of compelling and lively virtual worlds. Their motion is influenced by complex behaviors, and the production of expressive character animation heavily relies upon animators’ skills, training and knowledge. Ever since computers have been used in this creative process, researchers have strived to develop intuitive, interactive and precise tools and interfaces to facilitate animators’ work. In the last decade, deep learning methods have raised a lot of interest in the computer animation community, capitalizing on the increasing availability of real-life captured motion data to train generative models able to synthesize new movement. While the studied application allow for the generation of new realistic animation, the process unfortunately leaves little room for animators’ control. This thesis therefore proposes to use neural networks to learn the subtle complexities carried by motion data and to use the extracted information in the design of user-centric animation tools which fit in the existing animation production workflow. We propose two approaches allowing a user to edit a character’s pose at a given time frame, first by directly manipulating its skeleton, and second through higher-level pose parameters.