Self-supervised representation learning from demonstration

Given the growing demand for the application of robotics in an increasingly wide range of tasks and environments, robot learning as opposed to programming is getting more relevant by the day, since programming robots is tedious and usually only feasible in controlled domains. However, gathering data with robots is an expensive and time-consuming task, so the learning methods must cope with the limited amount of data available. When working with high dimensional perceptual data, learning low-dimensional, useful representations is a key aspect in dealing with the low-data setting. In this thesis, we develop a neural network architecture to learn perceptual representations from few human skill demonstrations in a self-supervised manner. The developed models take the sequentiality of the data and the low-data nature of the problem into account. These representations are used to learn perceptual goal models of the demonstrated skills. These models can monitor the learned skill executions and be used in reinforcement learning to generate reward signals, without explicit reward engineering. Our simulated and real robot evaluations with object manipulation skills show that the learned representations result in better goal models in terms of monitoring and reinforcement learning performance compared to generic dimensionality reduction methods. We further introduce transfer learning approaches in the context of learning from demonstration and show positive transfer between different objects for the same skill, between the same object for different skills under certain conditions, and also between different perceptual domains. Transferring knowledge as enabled by our modular neural architecture allows us to leverage existing data for continual learning into the future. Overall, we show that our proposed self-supervised representation learning architecture has the potential to improve learning from demonstration approaches with a perceptual component.