My research is driven by the goal of developing intelligent embodied agents that interact in the physical world. My research has primarily focused on the use of generative AI as an approach towards this goal. A major challenge in applying generative AI in this setting is the lack of available decision-making data and the necessity to generalize well to previously unseen situations. My work addresses this by constructing composable generative models using the idea of learning energy landscapes (EBMs) as a means to generalize beyond the narrow amount of data that is available, and some of my early work on EBMs led to the resurgence of diffusion models in 2020. EBMs enable compositional visual generation and compositional scene understanding. My work has further focused on how such compositional models can enable the synthesis of new trajectories in trajectory planning, enabling flexible adaptation to novel goals and rewards across both synthesized videos and on real robots. Finally, an energy optimization perspective on prediction enables us to combine the strengths of large pre-trained models together at prediction time, enabling both hierarchical planning and multimodal perception. More broadly, I am interested in constructing a decentralized generative architecture for decision-making, consisting of a society of different multimodal models, each with separate responsibilities such as 3D perception, memory, and auditory understanding, which jointly cooperate to make decisions in an environment. I am also interested in additional techniques to improve generative models such as reinforcement learning training as well as broader applications of my research in domains in science such as computational biology.