CAREER: Optimal Experimental Design through Contact: Towards Robots that Plan to Learn

Yale University, New Haven CT

Investigators

Abstract

The ability of humans to adapt in new environments and learn from a few interactions with their surroundings is long sought after in robotics. For example, when we step on ice, it only takes a few shuffles of our feet until we glide. Similarly, we are able to grasp and handle arbitrary objects with a few interactions. The closest robotics has achieved requires immense amounts of data, computation, and preexisting knowledge of the surroundings which pose significant barriers to obtain a glimpse of human-like capabilities. What if instead robots can plan for interactions that are beneficial for learning? Robots would have the potential to adapt to unseen and dynamically changing environments, broadening their utility in scenarios such as space exploration, deep ocean expeditions, and in humanitarian services like search and rescue. This Faculty Early Career Development (CAREER) grant seeks to develop these capabilities for robots to quickly adapt and learn new manipulation and locomotion skills by planning to learn through intentional interactions. Project activities include a synergistic educational and outreach plan in line with the PI’s goal of a well-rounded education for veteran and English as a Second Language (ESL) students in robotics. This plan includes a curriculum for integrating theoretical and practical courses in robotics, an educational program for literacy in science in a broad range of languages for ESL students, and undergraduate research opportunities for veterans. This project will support the PI’s goal of educating the next generation of capable roboticists. Motivated by how humans learn with just a few interactions, this project will advance how robots optimize and plan informative interactions for quickly learning locomotion and manipulation skills. The approach is centered on optimizing planned contact interactions that allow robots to take an active role in learning. The project plans an optimal experimental design formulation for reasoning about motion, contact interactions, and learning outcomes as a unifying optimal control problem. The effects of modeling choices in the planned formulation will be investigated. In addition, an online, model-predictive control (MPC) approach will be developed using repeatable dynamic learning primitives for real-time, deterministic, and reproducible learning behaviors. Furthermore, this project will demonstrate guarantees of reproducibility and certified learning as a result of robots influencing their learning outcomes through motion. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.