Character Control is the topic of my Master's Degree thesis; since I wanted to change from the usual rendering stuff I decided to focus on animation. The thesis's idea comes from "Near-optimal Character Animation with Continuous User Control" Treuille, A. Lee, Y. Popović, Z. ACM Transactions on Graphics 26(3).

The work is wide, thus it is difficult to explain every single part of it in details. I've put a link to my thesis in pdf format at the end of this page, so if you are interested, you can download it and give it a look.
The following list summarize most of the features present in the thesis while the next paragraph, which contains the abstract of my thesis, presents the idea behind the work.

• Skeleton Animation
• Motion Capture (acquisition, tracking, conversion from segments to transformation matrices, postprocessing)
• Animation Blending for skeleton animation
• Control System
• Reinforcement learning algorithms used to produce the near-optimal policy.
• Linear programming basis function approximation using ILOG CPlex

"The present work discusses theory and practice of a powerful animation method designed to generate walk animations for digital characters. Our system permits to interact with the animation, allowing the user to change at run-time several animation’s parameters, such as the motion direction or the gait style, influencing the final animation. Our work starts presenting the skeleton animation system and the motion capture system; then we explain how we can, thanks to these two techniques, generate a database of walk animation clips. The so obtained animations are provided to the animation system which links them together in a sequence. Linking is obtained generating a smooth transition from one clip to the next one through the use of the animation blending technique. Since we want to interact with the animation at run-time, the clip’s sequence is not given a priori, instead it is generated in real-time in respect to the user’s desires. To this aim we create a controller in charge of choosing the next clip in respect to the task that it is given, which can be simulating a walk along a line, or simulating a walk where motion direction and character’s orientation are required by the user. The controller leans on a selection policy in order to choose the next clip; in our work we propose two possible policies, a greedy one and a near-optimal one. The former goes through every animation contained in the database and evaluates the direct cost of adding the clip in exam to the sequence. The latter policy, instead, chooses the next clip evaluating an approximation of the optimal choice, which is obtained through the implementation of a reinforcement learning algorithm. The optimal choice estimates both the direct cost of choosing a clip as well as all the future costs that the system will pay for that choice. Unfortunately we can’t effectively implement the optimal policy, therefore we content ourselves with an approximation that leads to the nearoptimal policy. We lastly show how both these policies produce controllers capable of responding, in realtime, to the change of several animation parameters due to the user’s interaction as well as the environmental constraints."