Self-supervised and Transfer Learning for Adaptive Motion Control

非结构化环境中的运动控制是机器人技术中最难解决的问题之一。机器人可以熟练地完成日常和重复性的任务，但是今天没有一个机器人可以在没有人类繁琐的建模、设计和编程的情况下轻松地处理一项新的家务。然而，大自然在5亿年前就解决了这个挑战，对于人类的许多技能，我们并不知道一项技能是如何学习的，或者一项任务是如何完成的。虽然动物可以轻松应对复杂的环境，但运动控制仍然落后于大自然。在机器人研究中，基于模型的控制在机械手和有腿机器人中显示出了良好的效果。也就是说，如果我们有系统中每个刚体的运动学（即位置、速度、形状、尺寸等）和动力学（即质量、质心和惯性张量等）的完美信息，我们就可以找到实现期望运动所需的控制力矩。然而，这种技术依赖于运动学和动力学模型的准确性。虽然数据驱动的方法已被应用于解决重要的任务，但这种方法需要大量的人力劳动，并且受到感官噪音的影响。我研究的长期目标是开发鲁棒的机器人控制算法，这些算法在不同的场景下是自适应的和通用的。在过去的十年里，新开发的监督学习方法已经解决了许多需要找到一些输入和输出之间的映射的问题。在本提案中，我们的目标是通过新开发的机器学习技术来增强基于模型的运动控制。我们要讨论的一个例子是机器人抓取问题。具体来说，我们的短期目标分为以下三个主题：主题1：运动学和动力学的估计人类是将运动适应不同场景的专家。为了使机器人能够很好地工作，控制器需要了解物体的运动学和动力学模型。为此，第一个主题侧重于无先验知识的运动学和动力学估计。主题2：自我监督学习动觉教学，即手动移动机器人并记录运动，需要大量人力劳动。本主题关注自监督学习方法；也就是说，让机器人对不同的场景进行采样，并自己创建一个监督学习数据集。我们将努力设计一个标准，增加我们可以从数据集中获得的信息。主题3：模拟到真实的迁移学习本主题采用模拟到真实的迁移学习方法；也就是说，运动学和动力学模型将从部署在真实机器人平台上的仿真中收集的数据进行训练（离线）。通过这样做，我们可以减轻由感官噪声引起的问题，并确保计算可以实时完成。

Problem Motion-control in an unstructured environment is one of the hardest unsolved problems in robotics. Robots are skilled in routine and repetitive tasks, but no robot today can easily handle a novel household task without tedious modelling, designing, and programming by a human. Yet, nature solved this challenge half a billion years ago, and for so many human skills, we are not conscious of how a skill is learned or how a task is achieved. While animals handle complex environments with ease, motion-control is still lagging behind nature. Current state-of-the-art In robotics research, model-based control has shown promising results in manipulators and legged robots. Namely, if we have the perfect information of the kinematics (i.e., position, velocity, shape, dimensionality, etc) and dynamics (i.e., mass, center-of-mass, and inertia tensor, etc) of each rigid body in the system, we can find the control torques needed in order to achieve the desired motion. However, this technique is dependent on the accuracy of the kinematics and the dynamic model. While the data-driven approach has been applied to solve non-trivial tasks, this approach requires lots of human labour and suffers from sensory noises. Objectives The long term objective of my research is to develop robust robot control algorithms that are adaptive and versatile in different scenarios. Over the past decade, the newly developed supervised learning approach has been solving many problems that require finding the mapping between some inputs and outputs. In this proposal, we aim to enhance model-based motion control by the newly developed machine learning techniques. An example scenario that we will be targeting is robot grasping problem. Specifically, our short-term objectives are categorized into the following three themes: Theme 1: Estimation of Kinematics and Dynamics Humans are experts in adapting motion into different scenarios. In order for robots to perform well, the controller needs to have the knowledge of the kinematics and the dynamics model of the objects. For this, the first theme focuses on the estimation of kinematics and dynamics without prior knowledge. Theme 2: Self-supervised Learning Kinesthetic teaching, i.e., to manually move the robots and record the motion, requires a large amount of human labour. This theme focuses on the self-supervised learning approach; namely, let the robot samples different scenarios and creates a supervised learning dataset by itself. The effort will be on designing the criterion that increases the information we can gain from the dataset. Theme 3: Simulation-to-real Transfer Learning This theme takes a simulation-to-real transfer learning approach; namely, the models for kinematics and dynamics will be trained (offline) from data collected in simulation deployed on a real robotic platform. By doing so, we can alleviate the problems arisen from sensory noise and ensures the computation can be done in real-time.