Roy Fox Receives NSF CAREER Award
Roy Fox, assistant professor of computer science at UC Irvine Donald Bren School of Information and Computer Sciences (ICS), has received the National Science Foundation (NSF) CAREER Award. The award supports early-career faculty who have the potential to serve as academic role models in research and education and to lead advances in the mission of their department or organization. The award includes $655,580 over five years for Fox’s research on Modular Shared Robot Learning.
Fox is also the founder of the Intelligent Dynamics Lab (indylab) at ICS. His research interests include theory and applications of control learning — reinforcement learning, control theory, information theory, and robotics. His current research focuses on structure, exploration, and optimization in deep control learning of virtual and physical agents.
Fox sheds light on his research in robotics and the impact of the NSF CAREER Award here.
How would you describe your research?
My research is about how robots learn to do useful things, and in particular about who gets to teach them. In my group, the Intelligent Dynamics Lab, we work toward robots that are generalized in what they can do but personalized in how — learning machines that are competent across many tasks and environments, yet adapt to the particular people they serve. Robots are becoming increasingly capable of interacting autonomously with complex environments, but the knowledge of what a robot should do, and how a given person wants it done, lives with the user rather than with the engineers who built the robot. Today a user can teach a robot by demonstration, teleoperating it through a task so the robot can imitate what it saw. That is often a more intuitive interface than even language, but it places a heavy burden on the user, who may need to provide many demonstrations before the robot’s control policy generalizes to situations it has not encountered.
This NSF CAREER Award asks whether we can bring many more inexperienced users into directing the behavior of their own robots, through personalized continual learning built on the synergy of two conceptual pillars: shared learning and modular control. Shared learning means that knowledge gained in learning one robot’s task is reused in learning another, across different tasks, environments, robot bodies, and user preferences. Modular control means representing a robot’s policy not as one monolithic model but as a neuro-symbolic hierarchy of skills, in which lower-level skills such as picking up an object are the building blocks invoked by higher-level skills such as tidying a room, much like procedures in a software program. This modularity is what allows sharing to be fine-grained and practical at scale, letting us mix and match, skill by skill, what can transfer intact, what should be adapted, and what must be relearned from scratch. The three research thrusts — learning from imperfect and conflicting demonstrations, sharing skills across both similar and very different robots, and automatically discovering skill structure — come together in SkillHub, a robot learning system and shared skill library that will support both my research and educational activities.
What real-world impact does it have?
The prevailing path toward generalist robots today follows the one for AI more broadly: to train very large models centrally, on massive datasets, and deploy the same ones for all tasks and users. This project is one part of a complementary approach in which the users themselves participate in directing their robots, an approach I find scientifically interesting, technologically promising, and socially desirable. Scientifically, we would like to understand the capabilities and the limitations of algorithms that can share learning at scale among inexpert users. Technologically, this enables data collection at a scale far beyond what is available today and likely necessary for real-world robotic applications. Socially, democratizing robot learning improves how well these systems actually fit their purpose, and it preserves the autonomy of the people who live and work alongside them. Robots that are genuinely practical in the world will not only be trained before they ship; they will keep learning in the homes, labs, and workshops where they are used and be mindful of their human cohabitants.
The scientific questions here reach well beyond robotics. Structure and modularity matter almost universally, whenever a complex problem is composed of simpler parts. Understanding how knowledge accumulated on simple problems can guide the solution of complex ones, and how diverse learners can benefit from sharing that knowledge, has implications across machine learning more broadly, and for structured sequential decision-making problems as varied as business process automation, interactive theorem proving, virtual personal assistance, and hardware and software verification.
What does winning the NSF CAREER Award mean to you?
Robot learning is approaching an inflection point. The methods, the hardware, and the data are converging, and within the span of this award I expect robots that learn from the people they serve to move from research demonstrations into everyday settings, at a scale that will reshape how work gets done. This highly visible award places my lab and our institute at the leading edge of that tidal wave, and it gives us all a voice within it: not only larger models trained centrally, but robots that ordinary people can teach. That is the kind of frontier that attracts strong students, sustained collaborations, funding organizations, and industry partners who want to build an inspiring future together, and I am overjoyed for the opportunity to do work with such impact.