CPS:Medium:Collaborative Research: Safe Learning in Co-robots--Theory, Experiments and Education

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

This award will support fundamental research on scalable and safe collaborative cyber physical systems. Specifically, this project will address the question of how to balance adaptability with safety for large human-robot teams. While pre-programmed robots can perform well in a perfectly known and unchanging workspace, accomplishing complex tasks under real-world conditions requires the ability to adapt to unexpected circumstances. This adaptability can be provided using artificial intelligence or machine learning techniques, but the resulting robot behavior becomes less predictable, which in turn makes it difficult to guarantee safe human-robot interactions. This project addresses the challenge of guaranteed safety by using an innovative integration of machine learning with techniques from control theory and nonlinear dynamics. In contrast to current approaches, the methods studied here will be scalable, that is, they will remain practical to implement even when the number of interacting humans and robots become large. Many important applications can benefit from the results of this project, including disaster relief, rescue missions, homeland security, and assisted healthcare. This project will also develop a teaching platform for collaborative human-robot engineering that will be used to teach collaborative robotics in large classes, and to help broaden the participation of underrepresented groups in research. Robots that collaborate with human partners in a shared physical workspace are called co-robots or cobots. This research will address the fundamental challenge of safety and performance guarantees as collaborative human-robot cyber physical systems move from model-driven control approaches to data-driven methods. In particular, the project will focus on data-rich iterative tasks performed by groups of humans and robots, and dynamically challenging tasks where human-robot and robot-robot interaction forces are complex to model. Statistical learning theory will be merged with predictive control theory using a mix of physics-based and data-driven models in the learning process. In the co-robot cyber physical systems under study, robot and human models will be updated in real-time using data feeds. Within each robot such models will be used by a predictive controller to forecast robot motion and human interaction, and to take corresponding safe and collaborative actions. The new theory resulting from this project will provide statistically rigorous guarantees of performance improvement and safety during the learning process. 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.