University of Twente Student Theses


Bayesian learning-based impedance control of an aerial robot for writing

Avvari, V.S.Y. (2022) Bayesian learning-based impedance control of an aerial robot for writing.

[img] PDF
Abstract:In recent years there has been an exploding interest in extending the current applications of multirotor UAVs to those that require aerial physical interactions, such as contact-based inspection, aerial writing, and tool handling on hard-to-reach surfaces. Impedance control is a widely used interaction-control technique for aerial and ground robots. To achieve consistent performance during the interaction tasks, an a-priori knowledge of the environment parameters is needed to adjust the impedance controller parameters accordingly. For the task of aerial writing on unknown surfaces, this unknown knowledge of the environment makes it challenging to achieve a consistent outcome for the interaction tasks. Therefore, a novel method of finding optimal impedance controller parameters for interacting unknown surfaces is proposed in this thesis. The proposed method has two parts - use of pre-trained neural networks to predict the optimal controller parameters and generation of a custom dataset to train the neural networks. The work of this thesis is on the latter and involves a framework based on Bayesian Optimization (BO) to find the optimal parameters of an impedance-controlled aerial robot. Bayesian optimization is an iterative method of finding the maximum(or minimum) of an expensive black-box target function. A novel reward function is designed that depicts the accuracy and smoothness of the writing task by the aerial robot. Bayesian optimization is used to find the maximum of the target function formed by the impedance controller parameters and the rewards generated by the reward function. The corresponding controller parameters at the maximum reward are considered as optimal parameters. This is backed up by several simulations showing better accuracy in writing tasks with the optimal parameters predicted by the Bayesian optimization. Since the real-world robotic experiments are extremely dangerous and costly, a virtual aerial robot with same specifications as that of a real-world robot is created in the Gazebo simulator. Using the techniques of Sim2Real transfer learning such as domain randomization, various simulation scenarios are created by varying the parameters of virtual environment and noise added to the simulation. Bayesian optimization is used for every such simulation scenario and the optimal controller parameters are collected to form the custom dataset. In future, this dataset will be used for training the neural networks so that it can predict the required optimal controller parameters when the real-world aerial robot performs a similar writing task on an unknown surface.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Programme:Electrical Engineering MSc (60353)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page