Fast Kinodynamic Planning on the Constraint Manifold with Deep Neural Networks

Motion planning is a mature area of research in roboticswithmanywell-establishedmethodsbasedonoptimization or sampling the state space, suitable for solving kinematic motion planning. However, when dynamic motions under constraints are needed and computation time is limited, fast kinodynamic plan- ning on the constraint manifold is indispensable. In recent years, learning-based solutions have become alternatives to classical ap- proaches, but they still lack comprehensive handling of complex constraints, such as planning on a lower dimensional manifold of the task space while considering the robot’s dynamics. This article introduces a novel learning-to-plan framework that exploits the concept of constraint manifold, including dynamics, and neural planning methods. Our approach generates plans satisfying an arbitrary set of constraints and computes them in a short constant time, namely the inference time of a neural network. This allows the robot to plan and replan reactively, making our approach suitable for dynamic environments. We validate our approach on two sim- ulated tasks and in a demanding real-world scenario, where we use a Kuka LBR Iiwa 14 robotic arm to perform the hitting movement in robotic air hockey. Manuscript received 7 May 2023; revised 9 October 2023; accepted 15 October 2023. Date of publication 23 October 2023; date of current version 15 December 2023. This paper was recommended for publication by Associate Editor E. De Momi and Editor N. Amato upon evaluation of the reviewers’ comments. The work of Piotr Skrzypczy ́nski was supported by TAILOR, a project funded by EU Horizon 2020 under Grant 952215. The work of Krzysztof Walas was supported by REMODEL, a project funded by EU Horizon 2020 under Grant 870133. This work was supported in part by the CSTT fund from Huawei Tech R&D (U.K.), in part by the China Scholarship Council under Grant 201908080039, in part by the German Federal Ministry of Education and Research (BMBF) within a subproject “Modeling and exploration of the operational area, design of the AI assistance as well as legal aspects of the use of technology” of the collaborative KIARA Project under Grant 13N16274. The work of Piotr Kicki was supported in part by the Polish National Agency for Academic Exchange (NAWA) under the STER programme “Towards Interna- tionalization of Poznan University of Technology Doctoral School” (2022-2024) and in part by the PUT under Grant 0214/SBAD/0235. (Corresponding author: Piotr Kicki.) Piotr Kicki, Krzysztof Walas, and Piotr Skrzypczy ́nski are with the Insti- tute of Robotics and Machine Intelligence, Poznan University of Technology, 60-965 Poznan, Poland (e-mail: piotr.kicki@put.poznan.pl; krzysztof.walas@ put.poznan.pl; piotr.skrzypczynski@put.poznan.pl). Puze Liu and Davide Tateo are with the Department of Computer Sci- ence, Technische Universität Darmstadt, 64289 Darmstadt, Germany (e-mail: puze@robot-learning.de; davide.tateo@tu-darmstadt.de). Haitham Bou-Ammar is with the Huawei R&D London, CB4 0WG Cam- bridge, U.K. (e-mail: haitham.ammar@huawei.com). Jan Peters is with the Department of Computer Science, Technische Univer- sitätDarmstadt,64289Darmstadt,Germany,alsowiththeResearchDepartment: Systems AI for Robot Learning, German Research Center for AI (DFKI), 67663 Kaiserslautern, Germany, and also with the Hessian, 64293 Darmstadt, Germany (e-mail: jan.peters@tu-darmstadt.de). This article has supplementary material provided by the authors and color versions of one or more figures available at https://doi.org/10.1109/TRO. 2023.3326922. Digital Object Identifier 10.1109/TRO.2023.3326922