A knowledge-based planning framework for smart and autonomous manipulation robots

Manipulation planning in human environments is one of the challenging areas in robotics research. It is focused on making the robot capable of performing complex manipulation tasks, which requires manipulation planning capabilities in cluttered and unstructured environments. These capabilities need, on the one hand, a rich semantic description of the scene and knowledge about the manipulation actions, and on the other, a smart combination of task and motion planning levels. A method, called K-TMP, is proposed here where: a) knowledge is coded as an ontology framework with information about objects, robots, sensors and actions, the workspace and the context, as well as an inference mechanism for reasoning over this knowledge; b) planning is done with an heuristic task planner based on the Feed Forward method that uses a physicsbase motion planner to guide the search and find a feasible sequence of actions to perform the task. Knowledge is used at motion level to evaluate the potential interactions between the robot and the objects, resulting in robust paths in cluttered scenarios with the robot possibly interacting with some objects. At task level knowledge is used to take into consideration all the constraints affecting the actions, making the planning efficient. A table-top example with a bi-manual robot is included, as well as a discussion on challenges and future works.