Task-Oriented Manipulation Planning: Teaching Robot Manipulators to Learn Trajectory Tasks

As robot manipulator applications are conducted in more complex tasks and unstructured environments, traditional manual programming cannot match the growing requirements. However, human experts usually know how to operate robot manipulators to complete tasks, but they do not know how to manually program the robot for automatically executing tasks. From a general point of view, a robot manipulation task is composed of a series of consecutive robot actions and environment states which we call it trajectory task. Imitation learning, an emerging and popular technique of robot behavior programming, is a good way to tackle this line of work but still needs robotic and machine learning skills. Moreover, the state-of-the-art methods, such as inverse reinforcement learning methods, curriculum learning methods, and behavior cloning methods are suffering from expensive data collection, intensive data labeling, target goal recognition, and sometimes combinations of those challenges. In order to solve these challenges, we propose a method that can teach robot manipulators a variety of tasks without too many robotic and machine learning skills required. Besides the drawbacks of the state-of-the-art methods, there are new challenges: noisy demonstration data, a limited number of demonstration episodes, and a low random exploring success rate. To tackle this problem, we disassembled it into three parts: demonstration episode evaluation, demonstration guided trajectory generation and utilizing vision sensors for trajectory generation. These three parts correspond to chapters 2, 3, and 4 which state the details of each challenge. From the results, our proposed method outperforms the state-of-the-art methods and can be applied to different tasks.