Placement generation and hybrid planning for robotic rearrangement on cluttered surfaces

Official URL: http://risc01.sabanciuniv.edu/record=b2315123_ (Table of contents)

Rearranging multiple moving objects across surfaces, e.g. from a table to kitchen shelves as it arises in the context of service robotics, is a challenging problem. The rearrangement problem consists of two subproblems: placement generation and rearrangement planning. Firstly, the collision-free goal poses of the objects to be moved need to be determined subject to the arbitrary geometries of the objects and the state of the surface that already includes movable objects (clutter) and immovable obstacles on it. Secondly, after the goal poses of all objects have been determined, a plan of physical actions must be computed to achieve these goal poses. Computation of such a rearrangement plan is difficult in that it necessitates not only high-level task planning, but also low-level feasibility checks to be integrated with this task plan to ensure that each step of the plan is collision-free. In this thesis, we propose a general solution to the rearrangement of multiple arbitrarily-shaped objects on a cluttered flat surface with multiple movable objects and obstacles. In particular, we introduce a novel method to solve the object placement problem, utilizing nested local searches guided by intelligent heuristics to efficiently perform multi-objective optimizations. The solutions computed by our method satisfy the collision-freeness constraint, and involves minimal movements of the clutter. Based on such a solution, we introduce a hybrid method to generate an optimal feasible rearrangement plan, by integrating ASP-based high-level task planning with low-level feasibility checks. Our hybrid planner is capable of solving challenging non-monotone rearrangement planning instances that cannot be solved by the existing geometric rearrangement approaches. The proposed algorithms have been systematically evaluated in terms of computational efficiency, solution quality, success rate, and scalability. Furthermore, several challenging benchmark instances have been introduced that demonstrate the capabilities of these methods. The real-life applicability of the proposed approaches have also been verified through physical implementation using a Baxter robot.