Task and motion planning: scaling up

Abstract

This master thesis contributes to provide planning capabilities to embodied agents so that they are able to decide what sequence of actions should be taken to solve a given complex problem, involving coupled reasoning both at symbolic and geometric level (i.e. combined task and motion planning). This work has investigated width-based search algorithms, and a new algorithm has been proposed which implements a novel lazy-interleaved approach to reduce the computational cost of the planning, without making any assumptions on the robot used or the environment objects. Note, then, that the proposed approach works completely online without needing any precomputation at all. This work effectively exploits domain knowledge provided via sketches to guide the search in problems with huge-combinatorial state-spaces. The proposed approach is able to work transparently with robotic continuous search state spaces with flexibility, using an adaptive sampling of the world to build the search space of the problem. Moreover, the problem dynamics are retrieved as a black box, so the developed planner is able to work directly with a simulator, and it does not need an explicit declaration of the action structure. The proposed approach has been validated in two problem families that illustrate the current challenges in combined task and motion planning (video demos showing plans calculated using the developed framework can be found in https://drive.google.com/drive/ folders/10goVJ8A86RGIGsbthdmak8L_mTdT6hby?usp=sharing). Furthermore, the proposed approach has been compared with state-of-the-art approaches, obtaining significantly better results. Besides, the proposed approach has been combined and integrated within the ROS environment in order to provide the highest level of standardization and compatibility with the maximum number of robotic systems currently available.