Learning to Collaborate: Toward Robust, Adaptive Policies for Human–Robot Teams

Abstract

Robotics, automation, and the use of Machine Learning (ML) algorithms have been steadily making progress. They have been adopted in various sectors, including manufacturing, education, healthcare, and transportation. Although intelligent algorithms behind text-based, image-based, and commerce platforms are very prominent and often cited as examples of progress, there exists a gap in applying these algorithms to robotics applications with humans in the loop. In addition to human acceptance, robotic systems require safety-critical interfaces with humans (e.g., self-driving technology and robotic-assisted living). There is also a need for robot-specific datasets to train these algorithms. Providing efficient ways to train algorithms and building intuitive and safe interfaces with humans can lead to increased adoption and trust between end-users and robotic systems. The paradigm of Human--Robot Collaboration (collaborative autonomy) has been one of the most promising approaches to gathering data from human users and incrementally building trust between the human and the machine. However, humans are not static agents. Algorithms working with humans must consider the dynamic nature of their interactions with human users. This creates an exciting and challenging opportunity to develop algorithms that learn from humans and adapt to the requirements of an evolving task. In this dissertation, we investigate how robots can be trained efficiently and robustly given the dynamic nature of humans. Concretely, we explore three key objectives: (1) developing algorithms that learn efficiently from limited human demonstration datasets, (2) developing decision-making policies for long-term interaction, and (3) developing robot policies that communicate the learning to humans. This research leverages existing methods and builds on them to present novel approaches for learning from, communicating with, and adapting to human users. Our results are agnostic to the application domain (e.g., healthcare or driving) and to the type of robot (e.g., robot arm vs. autonomous car). Our main contributions are: (1) a learning algorithm for efficiently learning from human teachers, (2) a foundational optimization framework for influencing human partners over long-term interactions, and (3) a game-theoretic approach to communicating robot learning to human partners. We provide algorithms and experimental results from evaluations in simulated and real environments that demonstrate the effectiveness of our proposed approaches.