Achievable push-manipulation for complex passive mobile objects using past experience

The majority of the methods proposed for the problem of push-manipulation planning and execution deal with objects that have quasi-static properties and primitive geometric shapes, yet they usually use complex physics modeling for the manipulated objects as well as the manipulator. We propose an experience-based approach, where the mobile robot experiments with pushable complex real world objects to observe and memorize their motion characteristics together with the associated uncertainty in response to its various pushing actions. Our approach uses this incrementally-built experience to construct push plans based solely on the objects' predicted future trajectories without a need for object-specific physics or contact modeling. We modify the RRT algorithm in such a way to use the recalled robot and object trajectories as building blocks to generate achievable and collision-free push plans that reliably transport the object to a desired 3 DoF pose. We test our method in a realistic 3D simulation environment as well as in a real-world setup, where a variety of pushable objects with freely rolling caster wheels need to be navigated among obstacles to reach their desired final poses. Our experiments demonstrate safe transportation and successful placement of the objects.