Post Thread

This Penn Engineering research team draws on combined expertise from physics, mathematics, and computer science to build low-data robot learning systems that eliminate the need for massive training datasets to pick up new skills. The team’s two core innovations include a skill framework that lets robots build complex long-sequence tasks from basic learned movements, and a Gaussian Graph workspace roadmap that stitches a small set of existing motion demonstrations together to help robots navigate untaught paths and adapt in real time if obstacles interrupt their route. The lab is currently pairing these systems with language models and adapting them for dual-arm, humanoid and mobile manipulator robots, with the end goal of creating flexible, adaptive robotic partners that can reliably operate in the messy, unpredictable unstructured spaces of everyday real life, and the work has earned top honors at several major global robotics conferences.