Selecting where a robot should position itself to execute a desired grasp remains a fundamental challenge in mobile manipulation. Existing approaches often rely on optimization or precomputed reachability maps, which can struggle to scale to complex robots and environments. This project investigates learning conditional distributions over feasible whole-body robot configurations given a target end-effector pose. Rather than producing a single inverse kinematics solution, the goal is to develop generative models that capture the multimodal structure of valid base and arm configurations while incorporating practical constraints such as visibility, manipulability, collision avoidance, and robustness for motion planning. We will explore diffusion, flow-matching, and other conditional generative modeling techniques across increasingly complex robotic platforms, ranging from fixed-arm systems to high-degree-of-freedom mobile manipulators. By generating diverse, high-quality configuration proposals in real time, we aim to improve the efficiency and reliability of robotic planning and manipulation in real-world environments.

I am participating in SuperUROP because I want to gain more experience in robotics research. In particular, I am interested in building models that can eventually move outside our laptops and into the real world, where they can interact with complex and dynamic environments. More broadly, I see this project as an opportunity to explore what long-term research could look like for me. I’m excited to deepen my understanding of generative model architectures for robotics, contribute to an active research effort, and hopefully work toward writing a paper by the end of this experience.