Learning compositional models of robot skills for task and motion planning

The objective of this work is to augment the basic abilities a robot by learning use sensorimotor primitives solve complex long-horizon manipulation problems. This requires flexible generative planning that can combine primitive in novel combinations and, thus, generalize across wide variety In order plan with actions, we must have models actions: under what circumstances will executing successfully achieve some particular effect world? We use, and develop improvements to, state-of-the-art methods for active sampling. Gaussian process constraints on skill effectiveness from small numbers expensive-to-collect training examples. addition, efficient adaptive sampling generating comprehensive diverse sequence continuous candidate control parameter values (such as pouring waypoints cup) during planning. These become end-effector goals traditional motion planners then full performs skill. By using conjunction, take advantage strengths each dynamic tasks. demonstrate our approach an integrated system, combining robotics newly learned task planner. evaluate both simulation real world through measuring quality selected actions. Finally, apply system simulated real-world