Multi-Modal Path Planning for Solving Abstract Robot Tasks

For intelligent robots to solve real-world tasks, the problem is not only to plan motion paths, but rather to plan for picking, pushing, sliding, and many other diverse manipulation actions in a complex world of movable objects. In this thesis, we present algorithms, which are able to plan for manipulation and follow the multi-modal nature induced by these actions. We extend basic sampling-based motion planning to integrate Diverse Action Manipulation (DAMA) [6], and show that based on the Rapidly-exploring Random Tree (RRT), we can then solve DAMA scenarios of various kinds. We present three DAMA solving algorithms, which are build upon one another. To show the generic approach of our software for solving abstract tasks with various robot platforms, we evaluate one challenging scenario for a two-dimensional mobile robot, and one even more difficult scenario for a joint robot with ten degrees of freedom in three-dimensional space. The latter scenario was also executed in a real environment to illustrate the feasibility of the whole process from modeling to planning, up to execution. Results reveal that for the second scenario, in 78% of the cases, even our non-hierarchical algorithm finds a solution in under 15 minutes, despite about 93% of the time being wasted on computations related to inverse kinematics and nearest neighbor search, which is another area of interest and can be separated from our work. The contribution of this thesis is to give an overview of sampling-based manipulation planning, and to provide software and helpful implementation details, but also directions for future investigation to efficiently solve DAMA problems with various robot platforms in an environment featuring multiple objects.