Abstract: Reinforcement learning for robotics has enabled remarkable progress toward embodied intelligence, particularly for tasks and environments represented during training. In parallel, recent results suggest that large generative models can compress broad regularities in data into compact representations that support prediction and action generation, while also revealing limitations under distribution shift and novel combinations of familiar concepts. Current methods are still missing a key element of truly intelligent behavior: reliable generalization to new situations and new combinations of familiar skills. This talk presents research on compositional methods for robot design and control, which can generate new behaviors from simpler components. I will begin with templates as task-level abstractions for complex dynamical systems, then discuss how templates can be composed across behaviors and robot morphologies. Through examples in legged and flapping robots, I will show how reduced-order models support analysis, design, real-time control, and modular combinations of model-based and learned components. I will conclude by comparing this approach with end-to-end reinforcement learning: while learned policies can be highly efficient at execution, reduced-order model-based controllers offer explicit task structure, constraint handling, and the possibility of generalizing through recombination. This motivates a research agenda in which reduced-order models and learning work together to construct new behaviors from reusable dynamical components.

The Dartmouth Autonomy Seminars explore how common principles of autonomy link fields such as robotics, economics, and cognition. It brings together academia and industry to discuss autonomous systems. More information can be found here: https://sites.dartmouth.edu/das/