Yuchen Cui

Research Abstract:

As robots gradually get deployed into human-centered environments such as hotels, hospitals and our homes, they are challenged with novel scenarios and unseen data. Interactive machine learning holds the promise of allowing these robots to learn from user-generated data after deployment. At the same time, recent advances in supervised learning have shown great promise of large-scale offline pre-training of deep models and fine-tuning with human guidance. However, it is non-trivial to apply this paradigm to robotics applications for reasons including a large domain gap between pre-training and downstream robotics tasks, and availability of expert guidance. My research work centers around how to adapt robots with human guidance. The core objective of my research is to build autonomous agents that can efficiently adapt to new environments and tasks by interacting with non-expert human teachers. Learning from interactions with non-expert teachers introduces algorithmic and methodological challenges that prevailing machine learning methods can yet address, such as learning from inattentive teachers that provide sparse and noisy feedback, or learning from an expressive teacher providing diverse forms of feedback including verbal cues, facial expressions, and gestures. At the same time, we do not want to burden the user with spending a lot of effort teaching the robot, defeating the purpose of using the robot to make their life easier. I take two main approaches towards addressing this problem, one is to develop active learning algorithms that help robots to actively acquire informative data; the other is to harness as much data as possible when passively learning, including learning from implicit human feedback, and interpreting human gesture instructions. My research strives to make robots learn and adapt like humans do: by actively seeking knowledge, benefiting from diverse sources of feedback, leveraging pre-existing knowledge, and adjusting on-the-fly based on real-time guidance.

Bio:

Yuchen works at the intersection of machine learning and human-robot interaction, particularly for applications in home robotics. Her focus is on enabling low-effort teaching for non-expert users, aiming to develop algorithms that enable robots to learn efficiently from human interactions. Yuchen is currently a postdoc at Stanford's ILIAD lab working with Dorsa Sadigh. Yuchen obtained her PhD from the University of Texas at Austin in 2021, and her advisor was Scott Niekum.