Lifelong Learning for Vision based AUV Control

项目主要目标：主要目的是设计一个低层次的控制器的自主水下航行器（AUV）使用学习框架。该算法应该是跨多个机器人的便携式的，控制器应该快速适应动态的任何变化（有效载荷的增加、电流干扰、对机器人的损坏、推进器效率衰减等）。研究项目：水下机器人的控制是一项困难的任务，主要是由于水下动力学的非线性。经典的控制方法，如PID或滑模控制器，通常以低速运行，需要大量的专业知识来调整不同的参数。参数通常适用于机器人的一个特定配置，每次机器人配置改变时都需要进行调整。为了克服这些限制，本研究将研究基于学习的方法，并尝试将它们与经典控制方法相结合。基于学习的方法允许使用在机器人操作期间收集的数据来改进控制器。将这些方法与经典的控制方法相结合，使我们能够加快学习过程，因为大多数最近的学习方法需要大量的数据和学习时间。

Principal goal for project: The main objective is to design a low-level controller for autonomous underwater vehicles (AUVs) using learning frameworks. The algorithm should be portable across multiple robots, The controller should adapt quickly to any change in the dynamics (increase of payload, current disturbances, damage to the robot, thruster efficiency decay, etc...)Research project: Control of AUVs is a difficult task mainly due to the non-linearity of the dynamics in the underwater domain. Classic control approaches, such as PIDs or sliding mode controller, usually operate at a low speed and require a lot of expert knowledge to tune the different parameters. The parameters usually work for one specific configuration of a robot and tuning is necessary every time the robot configuration changes.To overcome those limitations, this research will look at learning-based methods and will try to combine them with classic control methods. The learning-based methods allow using data collected during robot operation to improve the controller. Combining those methods with classic control approaches allow us to speed up the learning process as most of the recent learning methods require a lot of data and learning time.