Advanced Manipulation with Darrt Planner

In order for robots to work well in human environments they must be able to plan for and execute many different types of manipulation.

During her internship last spring at Willow Garage, Jenny Barry from MIT implemented an algorithm for planning with multiple types of manipulation on the PR2.  This algorithm takes as input a starting configuration of the robot and objects, and outputs a sequence of robot trajectories corresponding to different types of manipulation.  For example, the planner could return a plan that first moves the arm to pick up an object, then moves the base over to another table, and then moves the arm again to place the object on the table.

By planning for multiple primitives at once we are able to ensure that our initial actions do not preclude later actions.  For example, we can guarantee that the grasp used in a pick does not collide with the environment during the place.  We also can plan for types of manipulation during which the object is not rigidly attached to the robot, such as pushing, sliding or throwing.  Currently, Jenny has implemented Pick, Place, and Push on the PR2.

For more information visit here for more info on the darrt planner and J. Barry, K. Hsiao, L. Kaelbling, T. Lozano-Perez.  Manipulation with Multiple Action Types.  ISER 2012