Bi-level planning approaches have seen recent success in solving robotic planning problems. The outer level of this planning scheme consists of a high-level symbolic plan, for example, Pick up the block. Move to the shelf. Drop the block.” In the low-level plan, non-linear constraint optimization algorithms optimize a set of actions for the robot, conditioned on the high level plan. While powerful, these optimizers are very sensitive to random initialization, and are not able to generalize well to longer length plans. In this work, we plan to use learning based approaches to intelligently initialize these optimizers, which will not only improve the robustness of this bi-level planning approach but also its speed.”

From my prior research experiences, I’m interested in thinking about some of the more nuanced problems in AI (interpretability, robustness, generalizability) and applying those ideas to designing intelligent embodied systems. My goal from this project is to publish.