R7-[PlanSpace] - Details

Spatial Planning

 

The project R7-[PlanSpace] addresses the manipulation planning problem in a mobile manipulation context. Current planning systems define grasps and placements either by discretization or by predefined continuous grasping surfaces. Our approach does not require such precomputation, but generates poses during planning. This leads to an infinite branching factor during planning in such problems, e.g., when placing an object its exact position is determined by continuous coordinates. For this purpose, a classical symbolic planner is extended to allow for ungrounded task specifications and unknown branching factors.

Many real-world manipulation tasks explicitly require the interaction with objects. For example, the manipulator should pick up and place a box filled with objects instead of picking and placing each of the objects separately, as is done in current manipulation systems. Therefore, we extend the capabilities of our planning system to overcome the lack of taking the physical properties of objects into account.

We account for multiple rigid or deformable objects or fluids and their physical behavior. The integration of a physical simulation environment in the planning process makes it possible to rule out infeasible configurations such as turning the box with the objects upside down.

However, integration of physical simulation into the planning process inefficient in general. Thus, we generate general physical constraints for manipulated objects using the physical simulation environment. We integrate those in the planning process, leading to an efficient implementation.