When designing policies or controllers for autonomous embodied systems, factors such as safety and data efficiency become paramount. For both low-level control and high-level planning problems, the standard practice has been to leverage symbolic domain knowledge (e.g., the governing equations of motion for the system, or an automaton representation of the high-level states) to design structured models that have certifiable guarantees, good generalization, or both (\eg, \citet{propel}). An emerging research direction is to automatically learn or discover the structure of the symbolic knowledge (\eg, \citet{xu2018neural}), which can be viewed as an instance of neurosymbolic programming.

---

Selected Publications