Character Animation

Human motion plays a key role in the production of films, video games, virtual reality applications, and the control of humanoid robots. Unfortunately, it is hard to generate high quality human motion for character animation either manually or algorithmically. As a result, approaches based on motion capture data have become a central focus of character animation research in recent years. We observe three principal weaknesses in previous work using data-driven approaches for modelling human motion. First, basic balance behaviours and locomotion tasks are currently not well modelled. Second, the ability to produce high quality motion that is responsive to its environment is limited. Third, knowledge about human motor control is not well utilized. This thesis develops several techniques to generalize motion capture character animations to balance and respond. We focus on balance and locomotion tasks, with an emphasis on responding to disturbances, user interaction, and motor control integration. For this purpose, we investigate both kinematic and dynamic models. Kinematic models are intuitive and fast to construct, but have narrow generality, and thus require more data. A novel performance-driven animation interface to a motion database is developed, which allows a user to use foot pressure to control an avatar to balance in place, punch, kick, and step. We also present a virtual avatar that can respond to pushes, with the aid of a motion database of push responses. Consideration is given to dynamics using motion selection and adaption. Dynamic modelling using forward dynamics simulations requires solving difficult problems related to motor control, but permits wider generalization from given motion data. We first present a simple neuromuscular model that decomposes joint torques into feedforward and low-gain feedback components, and can deal with small perturbations that are assumed not to affect balance. To cope with large perturbations we develop explicit balance recovery strategies for a standing character that is pushed in any direction. Lastly, we present a simple continuous balance feedback mechanism that enables the control of a large variety of locomotion gaits for bipeds. Different locomotion tasks, including walking, running, and skipping, are constructed either manually or from motion capture examples. Feedforward torques can be learned from the feedback components, emulating a biological motor learning process that leads to more stable and natural motions with low gains. The results of this thesis demonstrate the potential of a new generation of more sophisticated kinematic and dynamic models of human motion.