A constrained instantaneous learning approach for aerial package delivery robots : onboard implementation and experimental results

• Main Authors: Mehndiratta, Mohit, Kayacan, Erdal
• Other Authors: School of Mechanical and Aerospace Engineering
• Format: Article
• Language: English
• Published: 2020
• Subjects: Engineering::Mechanical engineering::Robots; Engineering::Mechanical engineering::Mechatronics; Instantaneous Learning; Learning-based NMPC
• Online Access: https://hdl.handle.net/10356/141838
• Institution: Nanyang Technological University

Rather than utilizing a sophisticated robot which is trained—and tuned—for a scenario in a specific environment perfectly, most people are interested in seeing robots operating in various conditions where they have never been trained before. In accordance with the goal of utilizing aerial robots for daily operations in real application scenarios, an aerial robot must learn from its own experience and its interactions with the environment. This paper presents an instantaneous learning-based control approach for the precise trajectory tracking of a 3D-printed aerial robot which can adapt itself to the changing working conditions. Considering the fact that model-based controllers suffer from lack of modeling, parameter variations and disturbances in their working environment, we observe that the presented learning-based control method has a compelling ability to significantly reduce the tracking error under aforementioned uncertainties throughout the operation. Three case scenarios are considered: payload mass variations on an aerial robot for a package delivery problem, ground effect when the aerial robot is hovering/flying close to the ground, and wind-gust disturbances encountered in the outdoor environment. In each case study, parameter variations are learned using nonlinear moving horizon estimation (NMHE) method, and the estimated parameters are fed to the nonlinear model predictive controller (NMPC). Thanks to learning capability of the presented framework, the aerial robot can learn from its own experience, and react promptly—unlike iterative learning control which allows the system to improve tracking accuracy from repetition to repetition—to reduce the tracking error. Additionally, the fast C++ execution of NMPC and NMHE codes facilitates a complete onboard implementation of the proposed framework on a low-cost embedded processor.