Learning robot navigation and manipulation from demonstrations

点击翻译按钮获取中文摘要

Objectives:- To research and implement a Learning from demonstrations method of navigation for mobile robot to perform complex navigation tasks independent of their domain.- To research implement a Learning from demonstrations method of navigation for mobile manipulator robots to perform complex manipulation tasks as well as complex navigation tasks independent of a robot's domain.- To implement autonomy on a mobile manipulator machine, which is usually teleoperated by an operator, using the researched Learning from demonstrations methods.Humans teleoperate machines to perform mobile navigation and manipulation tasks. Current autonomous system approaches are domain specific. Therefore, human operators are still in charge of the movement of their robots.This research studies Learning from demonstrations (LfDs) so the same teleoperated machines can be transformed to perform autonomously.The proposed research involves taking quantitative control data from human demonstrations. While a robot is learning a task from a demonstration, it must decipher useful task information from the noise in the control data. The research extends to not only just being able to replay the demonstration, but to also to adapting the execution of the task according to variations within the robot's environment.LfDs methods for mobile robots and mobile manipulators already exist, however these methods do not generalise the task and depend on the robot's system dynamics being known. They also use sensors which are expensive such as LIDAR rather than camera sensors. The LfD methods I would research into, and implement on mobile robots and mobile manipulators, is inspired from the work into manipulator robots conducted by Dr. Amir Ghalamzan. However, remapping of the existing models for manipulator robots onto the mobile robots and mobile manipulators will not be enough to make these robots fully autonomous.I will be looking further into state-of-the-art deep learning methods so that the robots do not only mimic or imitate the demonstrated task. But be able to generate ways of emulating demonstrations and include those demonstrations when learning the task. The idea is to improve the execution of the task and be able to generalise to be independent of the robot's domain.The outcome of the research is to produce a computationally efficient and effective method of implementing autonomy on mobile machines. The human re-programmable nature of the LfDs for the robots will increase the level of robot adaptation, as robot experts will not be required to continually reprogram the robots.