Title: Personalized, Safe, and Interactive Robot Programming via Human Demonstrations

Date: Monday, April 22nd, 2024
Time: 10:00 AM - 12:00 PM EST

Location: TSRB Room 132

Virtual Link: Zoom

Letian Chen

Computer Science PhD Student
School of Interactive Computing
Georgia Institute of Technology

Committee:

Dr. Matthew Gombolay (Advisor) - School of Interactive Computing, Georgia Institute of Technology
Dr. Sonia Chernova - School of Interactive Computing, Georgia Institute of Technology
Dr. Harish Ravichandar - School of Interactive Computing, Georgia Institute of Technology
Dr. Benjamin Eysenbach - Department of Computer Science, Princeton University
Dr. Scott Niekum - College of Information and Computer Sciences, University of Massachusetts Amherst

Abstract:

The increasing capability of robots and machine learning algorithms shed light on the future where robots can be deployed ubiquitously. Yet, current robot learning algorithms require robotic and programming expertise, limiting the functionality users can gain from robots. Learning from Demonstration (LfD) techniques seek to democratize robot learning by empowering end-users the ability to teach robots new skills. However, most prior work overlooks several key factors for LfD algorithms to succeed in the hand of end-users, e.g., assuming humans accomplish tasks homogeneously and humans can provide high-quality demonstrations on their first try. In my thesis, I seek to fill the gap between LfD approaches and users by proposing LfD algorithms that allow robots to provide personalized and safe service after efficient, interactive training.

I first develop an algorithm that learns from heterogeneous human demonstrations in a federated, lifelong way by constructing and reusing prototypical policies to model diverse human preferences. I then extend the algorithm to the offline learning setting where the agent is not able to obtain more interactions with the environment beyond demonstration in high-stake tasks such as medical and Mars rover planning problems. These two algorithms allow robots to efficiently learn personalized policies from heterogeneous user demonstrations. Next, I consider the safety problem of the LfD policy, as the robot directly interacts with end-users. I create an algorithm that is the first to allow users define what they deem as safe, and shield learning from demonstration policy from ever taking unsafe actions. 

In my proposed work, I seek to create an interactive, multi-modal learning system between users and robots such that users can specify their intentions in either demonstrations or language, and can convey their intentions to robots iteratively in multiple rounds based on their observation of the robot behaviors, closing the loop in the robot learning from demonstration system.