Incremental Nonlinear Control and Reinforcement Learning for Safe Autonomous Airship Operation

• Grant number: 23/04266-7
• Support Opportunities: Scholarships abroad - Research
• Start date: April 15, 2024
• End date: August 29, 2024
• Field of knowledge: Engineering - Electrical Engineering - Industrial Electronics, Electronic Systems and Controls
• Principal Investigator: Ely Carneiro de Paiva
• Grantee: Ely Carneiro de Paiva
• Host Investigator: Giovanni Beltrame
• Host Institution: Faculdade de Engenharia Mecânica (FEM). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Institution abroad: École Polytechnique de Montréal, Canada

Abstract

This document presents a one-year proposal for a Sabbatical program of Prof. Dr. Ely Carneiro de Paiva (FEM-Unicamp) at École Polytechnique Montréal, Canada, aiming the development of a classical nonlinear flight controller as well as a Neural Network flight controller for the accurate positioning task of a heavy load airship.This proposal is related to the Thematic Project Fapesp-CNPq "INSAC-INCT in Applied Cooperative Autonomous Systems" (Fapesp 2014/50851-0 and CNPq 465755/2014-3), which is a 7-year multi-institutional and multi-disciplinary research project on Mobile Robotics, conducted by EESC-USP, joining more than 10 universities and research centers, including the support of the Instituto Superior Técnico (IST) of Lisbon, Portugal.Regarding the Canadian side, this cooperative program is inserted in our partners 3-year project named "ENVIA - Simulation environment for AI-driven vehicles", conducted by Polytechnique Montréal in partnership with Flying Whales and Thales companies, financed by CRIAQ - Consortium for Research and Innovation in Aerospace in Québec.This proposal has two main goals. The first one is to design an Incremental Dynamic Inversion Controller with guaranteed safe operation for the automatic control of an airship positioning over a ground target. The second one is to use this nonlinear controller to train a neural network through reinforcement learning (RL) aiming to increase the robustness of the system in accurate positioning against strong wind disturbances.This work will be developed at Polytechnique Montréal, Quebec, Canada, where we expect to benefit from the expertise of Prof. Dr. Giovanni Beltrame and his research group in the thematics of Perception and Artificial Intelligence (mainly in Reinforcement Learning). The proponent will have access to the unique infrastructure of the MIST Lab that has developed numerous technologies for multi-robot control, navigation and localization, including learning-based methodologies, with partners in industry and American universities, like MIT and Stanford.Both research groups will benefit from this cooperation project. Polytechnique can benefit from our 20-years experience in modelling, simulation, guidance and control of airships using classical control, while we can take advantage of their knowledge in Perception and Artificial Intelligence for further introducing machine learning - a theme on the frontier of science nowadays - in our mobile robotic projects. It is worthy to say that we have no researcher/professor with expertise in Intelligent Control at our faculty (FEM-Unicamp) at this moment.The results obtained will not be restricted to the robotic airship case. They shall be extended to other kinds of UAVs like the "drones" or quadcopters (theme of investigation of our partners in USP and UFMG of the Thematic Project), that are able to make hovering flight over a target in the ground, and are sensible to wind and gusts perturbations. The development of this high-accurate positioning controller may help to strength other important researches on the field of Aerial Robotics in Brasil.