Learning Task Models for Robotic Manipulation of Nonrigid Objects
Langsfeld, Joshua D.
Gupta, Satyandra K
MetadataShow full item record
As robots become more prevalent in smaller manufacturing and maintenance settings, it will become important to enable them to learn new tasks quickly without explicit programming by a human. One particularly challenging domain in robot learning is handling nonrigid objects and materials such as fluids and easily deformable parts and tools. The complexities of modeling nonrigid systems make it infeasible in general for a robot to plan its actions to perform a task by simulating their behavior, requiring an ability to learn an unknown model through experience. This experience can be gained both from a human demonstrating the way to perform a task and the robot itself performing task attempts to incrementally improve its model. Over time, as more experience is acquired, the robot should eventually obtain a model that allows it to perform the task when faced with new variations, generalizing its past experience. This dissertation explores this problem in the context of two robot tasks: pouring a specific volume of fluid into a moving container, and cleaning stains off of compliant objects. First, an approach is presented to learn the parameters of the pouring task by observing human demonstrations. The model learned from the demonstrations can then be exploited to learn how to pour new volumes with minimal extra learning effort by the robot. Second, this same task is used in development of a general approach for autonomous learning. Here, the robot takes a small set of random samples from the parameter space to build an initial task model and selects new parameters to test by building many local linear models. As more data is acquired, the robot's task performance improves substantially and it is able to very quickly find solutions to new task variations. Then another approach is shown that uses demonstrations to estimate a cost function for performing the task. This enables the robot to also learn strategy elements from how humans perform tasks. Finally, two approaches are discussed to learn the deformation model of a compliant part. A bimanual setup with two robot arms is used to hold and clean the part and the model is used to optimize the plans for both arms to reduce cleaning time and deformations. The first approach shows a black-box learning method to directly predict the part deformation when a known force is applied. The second uses a finite-element structure to represent the part, and learns the model by updating the stiffness parameters. When given a new part, the system only needs a few trials to improve quickly enough to clean new stains efficiently by predicting how much the part will deform under cleaning force.