We develop algorithms for robot learning, safe and efficient human-robot interaction and multi-agent systems.
Our mission is to equip robots, or more generally agents powered with artificial intelligence (AI), with the capabilities that will enable them to intelligently learn, align with, adapt to, and influence the humans and other AI agents.
We take a two-step approach to this problem.
First, machine learning techniques that we develop enable robots to model the behaviors and goals of the other agents by leveraging different forms of information they leak or explicitly provide.
Second, these robots interact with the others to achieve online adaptation by leveraging the learned behaviors and goals while making sure this adaptation is beneficial and sustainable.
Robot Learning
We develop robot learning algorithms that leverage various forms of data.
We are interested in developing machine learning techniques that enable AI agents to model the behaviors and goals of humans and other agents by leveraging different forms of information.
We specifically focus on learning from humans via multiple modes of information sources, including explicit forms such as human demonstrations and comparisons, and more implicit forms such as human gaze and gestures, to equip AI agents and robots with the capability to understand and align with humans' goals and preferences.
We also emphasize the importance of data-efficiency and expressiveness in learning and capturing humans' behaviors accurately.
We work on projects such as explainable reinforcement and imitation learning, modeling a belief over human preferences to tune it with their feedback, and developing active learning algorithms for data-efficient learning from comparisons.
Human-Robot Interaction
An application of our human-robot interaction work is personalizing exoskeletons.
We aim to improve the cooperation between humans and robots by equipping robots with the capability of using multiple modes of information sources to understand, align with, and adapt human goals and preferences.
For this purpose, we computationally model human behavior using techniques from machine learning, behavioral economics, cognitive science, and game theory.
Using the above-mentioned robot learning techniques, we have made strides in personalizing robots for various applications, such as home robots or lower-body exoskeletons for people with paralysis.
We also focus on developing partner-aware algorithms that allow AI systems to make better decisions in human-robot collaboration tasks by predicting humans' actions.
Multi-Agent Systems
We study multi-agent systems ranging from two-player games to systems with thousands of robots and humans.
Our work on multi-agent systems involves studying game-theoretic settings, and ensuring safety and efficiency in complex multi-agent environments.
By incorporating learning and adaptation strategies, we aim to create AI agents that can predict and coordinate with the other agents in the environment.
We have demonstrated the benefits of partner-aware agents in achieving better equilibria in canonical games and ensuring safety in multi-agent Markov decision processes.
We also investigate traffic optimization, showing that autonomous vehicles predicting human drivers' routing choices can help solve congestion problems and optimize pricing for ride-hailing or public transportation services.
