Building Adaptable Generalist Robots

  Over the past decade, advancements in deep robot learning have enabled  robots to acquire remarkable capabilities. However, these robots often strug gle to generalize to new, unseen tasks, highlighting the need for the devel opment of generalist robots. While existing research primarily focuses on  enhancing generalization through large-scale pre-training—providing robots  with vast datasets and extensive parameters and treating generalization as a  naturally emerging trait—this approach does not fully address the complex ities of the real world. The real world presents an infinite array of tasks,  many of which extend beyond the training scenarios previously encountered  by these robots. For example, in healthcare, robots must manage the partial  observability resulting from the diverse latent intents of patients, which are  not to be covered in the dataset. Similarly, autonomous vehicles must navi gate unpredictable traffic, weather, and road conditions, which may go beyond  the training data.  

This thesis contends that, alongside scalability, a strong adaptation capa bility is crucial for improving generalization in real-world applications. It  explores strategies for building robots that can adapt effectively at the time  of deployment, with a focus on data efficiency, parameter efficiency, and ro bustness. The study delves into various adaptive learning methods, including  in-context robot learning that conditions on a limited number of demonstra tions, unsupervised continual reinforcement learning that uncovers the struc ture of robot tasks, and the use of large foundation models for building embod ied agents. These methodologies demonstrate significant potential, enabling  robots to acquire new motor skills across diverse applications and solve com plex, long-horizon physical puzzles through creative uses of tools.