Building General Purpose Robots

Science fiction is always ahead of its technology contemporaries. In Dr. Asimov’s novelette \textit{The Bicentennial Man}, a housekeeper robot named ``Andrew'' was set to be introduced to the Martin family in 2005. But until now, massive applications of ``Andrew'' like general-purpose robots still remain absent. Generally, two aspects of challenges exist here: the environment and the task. With recent advances in deep learning, the challenge of the unstructured-environment has been hereto, perhaps well addressed due to the representation power of deep neural networks. But due to the complexity of robotic tasks, the task specification problem still lacks a satisfactory solution. Human demonstrations, compared to way-point based robot teaching and reward/cost based learning methods, provide a more intuitive interface for general purpose task programming. Moreover, learning directly by watching human demonstrations without either tedious kinesthetic teaching or teleoperation, is even more convenient for task teaching. However, just as the ``no free lunch'' theorem suggests, more convenience means more challenges. Generally, our trouble comes from the “correspondence problem” suggesting that humans and robots perform even the same task in fundamentally different ways both perceptually and physically. Perceptually, the observation space of human and robot could be different. For example, the camera used to record human demo videos can be different from the on-board sensor of a robot when executing the task. Physically, humans and robots have different action spaces. When collecting samples, we have no access to robot actions. So the strong task clue based on state action pairs in approaches like behavior cloning (BC), inverse reinforcement learning (IRL) or meta-learning, is no longer available. These extra difficulties make learning a controller or policy even more challenging in that massive robot-environment interactions are commonly required, which can be cost inhibitive in real-world applications. This learning paradigm aligns with the human cognitive process in peer learning and observational learning which involves first understanding the task by what has been observed before attempting any motor actions. So a global information of task definition is In this thesis, we tackle the above challenges by learning a geometry structured latent representation no robot actions are collected Moreover, since we only record video frames, there are no robot actions collected, Rethinking the ``correspondence problem’’, we can infer that only the information of task definition should be transferred from human demonstrator to robot imitator. Empirically, this aligns with the human cognitive process in peer learning and observational learning which involves first understanding the task by what has been observed before attempting any motor actions. Furthermore, suppose we can directly encode task definition from an observed image, then the task definition itself should also transfer across different environmental settings or categorical objects / tools. This gives us a hint to build generalizable robot task learning. Likewise, this is also similar to human performance in that humans can seize generalizable task definitions simply by watching others performing the same task under various task settings. This thesis concerns four questions: (1) how should we encode the task definition directly from an image? (2) how should we learn it from human demonstration videos? (3) how will the learned task definition encoding relate to a control or policy? (4) how should the task learning be generalizable? In this thesis, we introduce geometry as a structured priori to derive solutions. Next we will explain in more detail. the prior knowledge which structurizes the task representation learning from human demonstration videos. Then how should we extract task definitions directly from an image--- what are the considerations and what is the neural network architecture? How does the representation of task definition relate to robot controller design or policy learning? How should the task learning be generalizable? There could be many possible ways to design a task definition encoder, one simple example is a neural network that takes an input image and outputs a vector encoding the task definition. But in practice, such simple parameterization does not help addressing the tedious controller/policy learning problem. In this thesis, we introduce geometry as the prior knowledge which structurizes the task representation learning from human demonstration videos. We show that such geometry structured task representation enables the design of efficient controllers and the generalization to various objects and tools. There could be many ways to encode a task from images, for example, a neural network that takes an image input and simply outputs a task encoding vector, however such simple parameterization does not help addressing the tedious controller/policy learning problem when transferring the task definition to a robot. Can we bring some structured constraints to the neural network so that the learned task encoding enables an easier controller design or policy learning? Then say geometry is structured. 1. 2 Research Questions 1 什么是task representation learning problem，我们需要什么样的representatin，有什么性质？ 2 介绍geometry，介绍visual servoing。介绍为什么geometry可以有用。 3 总结，我们的approach是什么。解决了deep l的什么问题，解决了visua servoing的什么问题。 4. 分析引入geometry In this thesis, we introduce geometry into the neural network as a structured task representation. This motivation comes from decades of research proven effective techniques in the visual servoing literature that use geometric association constraints (Fig. 1, \eg, point-to-point, point-to-line, line-to-line) and their combinations to represent a task in a general way and to control the robot in an efficient way. For example, in our problem setting, the camera used to record human demonstration videos is different from the on-board one that guides robot actions. So the human demonstrator and robot imitator have two different observation spaces. Moreover, since we only record video frames, there are no robot actions collected, then the strong task definition clues using state action pairs in either behavior cloning (BC), inverse reinforcement learning (IRL) or meta-learning approaches, are no longer available. As a result, learning a robot controller or policy still requires massive data samples via an interactive learning environment in either a tediously designed simulator or the real robot. This can be cost inhibitive in real-world applications. \section{Research Questions} Formally we name the task definition encoding as the task representation, learning an encoder that extracts task definition from an observed image as the \textbf{task representation learning problem}. This research intersects with deep learning and traditional visual servoing. It combines the representation power of deep neural networks while maintaining advantages of visual servoing like data efficiency, good interpretability and almost no hardware wear-out during training. As a result, this research transforms the task specification problem in visual servoing to the problem of task representation learning by feeding several human demonstration videos. Our goal is to learn generalizable task representation from several human demonstration videos and We show that such geometry structured task representation enables the design of efficient controllers and the generalization to various objects and tools. Furthermore, we ask the question how to encode the task definition so that it enables an easier robot controller design or policy learning. Intuitively, we can directly formulate a neural network that takes input of an image and outputs a vector meaning the task definition. view on human demonstration towards encoding more task concepts rather than control we examine \textit{what to imitate} which allows a common shared knowledge transferring Recent advances in deep learning based approach have hereto well addressed the unstructured-environment challenge.