Embodied agents are increasingly expected to learn new skills from limited demonstrations and to operate continuously once deployed. Recent approaches have emphasized scale through larger video-generation world models, vision–language–action policies, and demonstration corpora. However, prediction, memory, and action are often studied independently, leaving open how representation structure supports learning and continuous operation across these capabilities. The thesis addresses this question through three complementary studies of prediction, memory, and closed-loop action.

For prediction, we study a world model that represents future manipulation motion as scene-level 3D point trajectories and adapts to a new robot from a few demonstrations. Its accompanying data pipeline converts heterogeneous human and robot videos into a shared geometric representation for learning across embodiments. For memory, we introduce a sequential evaluation protocol that examines whether embodied question-answering agents can reuse accumulated observations across successive queries, distinguishing memory persistence from useful knowledge accumulation. For action, we develop an industrial unpacking simulation environment to generate demonstrations and conduct a study of diffusion policies for precision fan insertion, examining how demonstration coverage and temporal context affect closed-loop reliability.

These studies highlight complementary roles for structure: geometric motion representations support transfer across embodiments, spatially organized memory supports reuse of past observations, and temporal context improves policy performance in the studied industrial setting. Together, they motivate designing representations and supervision around the information an agent needs to predict, remember, and act, and evaluating these choices during continued interaction with the environment. The thesis concludes with directions for connecting motion prediction, persistent scene memory, and policy learning within structured world models.