Events Calendar

PhD Thesis Proposal - Heramb Nemlekar

PhD Candidate: Heramb Nemlekar

Title: Efficiently Learning Human Preferences for Proactive Robot Assistance in Assembly Tasks

Date: 09/20/22 (Tuesday)
Time: 1:30 pm

Hybrid Presentation

Location: SAL 322
Zoom URL: https://usc.zoom.us/j/92429044874?pwd=ZExXK2dRamtWK1V0MUswOVVUdnJzZz09

Committee:
Gaurav Sukhatme
Heather Culbertson
Jyotirmoy Deshmukh
Satyandra K. Gupta
Stefanos Nikolaidis

Abstract:
I focus on enabling robots to proactively assist assembly workers by anticipating their future actions. Since each worker can have their own preferred way of performing an assembly, assistive robots must learn the individual preferences of their users to anticipate their actions accurately. Previous work in this field learns human preferences in the form of a policy or a reward function from user demonstrations in the given task. However, obtaining demonstrations can be tedious and time-consuming in actual assemblies. While recent approaches try to efficiently learn user preferences through actively generated queries instead of demonstrations, they do not leverage any prior knowledge of the users' preferences. In this work, I propose exploiting (1) similarities between preferences of different users in a given task and (2) similarities between different tasks performed by a given user to learn the user's preference efficiently. Based on our insight that different users can be grouped into a small set of dominant preferences, I present a novel approach for efficiently inferring the preferences of new users by matching their actions to a dominant preference cluster. For leveraging similarities between tasks, I propose learning user preferences as a function of task-agnostic features (like the mental and physical effort of actions) such that we can transfer their preferences from a canonical to an actual assembly task. I evaluate these approaches in user studies of real-world assembly tasks and show how each prior source of user preference can improve the accuracy of anticipating user actions. I also present a human-robot assembly study that shows how proactively assisting users by anticipating their actions can reduce human-idle time and improve user experience. Finally, I discuss our proposal for jointly leveraging similarities between users and tasks to accurately anticipate user actions.

Location: 322

WebCast Link: https://usc.zoom.us/j/92429044874?pwd=ZExXK2dRamtWK1V0MUswOVVUdnJzZz09

Audiences: Everyone Is Invited

Contact: Lizsl De Leon