Abstract: Solving long-horizon tasks involving manipulating multiple objects, such as cleaning a dirty room and preparing a meal, has long been a grand challenge in robotics. However, computationally intractable planning presents the biggest hurdle to achieving such tasks. In this talk, I will discuss task and motion planning (TAMP), a general planning framework that exhibits a bilevel structure: high-level task reasoning (i.e., determining which actions to take) and low-level motion reasoning (i.e., determining how to execute these actions) mutually provide complementary guidance. I will then introduce the fundamental challenges preventing the deployment of existing TAMP approaches to long-horizon problems in practice. Additionally, I will address these challenges by leveraging machine learning and multi-robot systems to enhance efficiency. Specifically, these ideas include: (1) learning when to halt planning, (2) investigating long-horizon dependencies using machine learning techniques, and (3) distributing tasks among multiple robots.

 

Short Bio: Yoonchang Sung is a postdoctoral fellow at UT Austin, working with Peter Stone, and was previously a postdoctoral associate at MIT CSAIL, working with Leslie Kaelbling and Tomás Lozano-Pérez. He completed his Ph.D. at Virginia Tech, advised by Pratap Tokekar. His research interests include robot planning and learning, task and motion planning, and multi-robot systems. He was selected as one of the RSS Pioneers in 2019. His work was nominated for the Best Cognitive Robotics Paper Award and received the Best Robocup Paper Award, both at IROS 2021