Harsha Vardhan Guda

Robots are unable to manipulate common household deformable objects, such as bags or clothes, with the same precision and speed as rigid objects. In transferring offline learned policies to the real world, the intrinsic complex dynamics of deformable objects cause a simulation-to-reality gap, which leads error to accumulate, making the direct application of such policies impractical. This thesis aims to bridge the gap in dexterity by developing algorithms to facilitate efficient learning of deformable object manipulation. This includes the study and development of new methodologies for learning parametrized cloth manipulation policies in simulation, to be used as baseline policies that are to be adapted and optimized through few real-robot executions. The resulting motion policies will be combined with proper model-based controllers and real-time feedback to increase manipulation speed and precision. These methods will be applied to tasks such as in-air manipulation of cloth, or also tasks requiring human-robot interaction such as assistive dressing.