Knowledge-Oriented Physics-Based Motion Planning for Grasping Under Uncertainty

Grasping an object in unstructured and uncertain environments is a challenging task, particularly when a collision-free trajectory does not exits. High-level knowledge and reasoning processes, as well as the allowing of interaction between objects, can enhance the planning efficiency in such environments. In this direction, this study proposes a knowledge-oriented physics-based motion planning approach for a hand-arm system that uses a high-level knowledge-based reasoning to partition the workspace into regions to both guide the planner and reason about the result of the dynamical interactions between rigid bodies. The proposed planner is a kinodynamic RRT that uses a region-biased state sampling strategy and a smart validity checker that takes into account uncertainty in the pose of the objects. Complex dynamical interactions along with possible physics-based constraints such as friction and gravity are handled by a physics engine that is used as the RRT state propagator. The proposal is validated for different scenarios in simulation and in a real environment using a 7-degree-of-freedom KUKA Lightweight robot equipped with a two-finger gripper. The results show a significant improvement in the success rate of the execution of the computed plan in the presence of object pose uncertainty.