Computational Acquisition of Robust Motor Skills

Humans have the ability to learn complex motor skills like ballet dancing, breakdancing, gymnastics and riding a bicycle. Ideally, these skills are robust to many conditions: they can be performed whether the day is windy or the ground is slippery. This thesis introduces computational tools for physics-based simulated characters to acquire robust motor skills. We show how to optimize control solutions to produce complex motor skills, such as breakdancing, handstands and hand walks. The cost function is specified in terms of intuitive motion properties (e.g., place the hands close to the ground, place the feet as high as possible, maximize angular momentum in some direction, etc.). Unlike previous work, our method does not require contacts to be pre-specified, which simplifies the synthesis process and allows tackling motions that were previously outside the scope of physics-based methods. Our approach is currently too slow for online applications that require control solutions for a range of tasks to be performed. We show that if we restrict the problem to specific movements, it is possible to design online controllers that achieve a range of tasks and that are robust to disturbances. We demonstrate this for a variety of rotational movements, including cartwheels, dives, flips and handsprings. The control solutions described so far are only valid for a small set of states of the character. We develop a method to systematically expand this set, thereby increasing the robustness of our control solutions and allowing for their sequencing. Once a skill is learned by a character with a particular body shape, we show how it can be transferred to a character with a different body shape. The method discovers surprisingly natural results, such as a heavier character avoiding losing balance by taking extra steps and stretching out its arms. For a given character, this opens up a large reservoir of skills that can be acquired.Ph.D