Simulating Human Motion for Object Manipulation
MetadataShow full item record
Performing object manipulation with full-body and dexterous hands is an essential human activity in everyday environment. Although problems involving object manipulation have been frequently studied by researchers in computer animation, virtual characters exhibiting the same level of versatility as their human counterparts have not been demonstrated. In this thesis, we identify three aspects of versatile manipulation skills based on the observation of human behaviors. First, humans are adept at organizing a set of independent tasks into consecutive and concurrent tasks that maximize efficiency -- putting the coffee mug on the roof of the car, opening the car door, and loading stuff into the car while holding a baby. Second, humans are good at utilizing different body parts for manipulation tasks -- using the elbow to push the door when both hands are occupied or using both palm and fingers to orient a cellphone in hand. Third, humans are skilled in manipulating a variety set of objects with different material properties -- the same manipulators (i.e. hands) are capable of folding laundry and hammering a nail. One existing challenge is to develop algorithms that can coordinate different object manipulation tasks. Moreover, it requires solving challenging planning and control problems to capture how humans deftly exploit different properties of different body parts. In addition, manipulation of deformable objects brings challenges in physical simulation and dynamic control due to complex collision phenomena and large number of degrees of freedom for the object. We use physics-based optimization and control techniques to simulate human motions for full-body and dexterous hand manipulation with versatility that humans exhibit, to address the abovementioned challenges. This thesis focuses on three research problems: (1) synthesize human activities involving concurrent full-body manipulation of multiple objects; (2) synthesize in-hand manipulation of a polygonal object with integrated control techniques for both palm and fingers; and (3) synthesize realistic dexterous manipulation of cloth from a simple description of the desired cloth motion. The algorithms presented in this thesis make a step further towards automatically creating animation for full-body and dexterous hand object manipulation with human versatility.