|
3:30 p.m.
Online presentation
For More Information:
Soheil Feizi
sfeizi@umd.edu
https://umd.zoom.us/j/95197245230?pwd=cDRlVWRVeXBHcURGQkptSHpIS0VGdz09
Computer Science Department Machine Learning Seminar Series
Learning Robot Policies from Non-Traditional Sources of Human Data
Dorsa Sadigh
Computer Science and Electrical Engineering
Stanford University
Zoom link: https://umd.zoom.us/j/95197245230?pwd=cDRlVWRVeXBHcURGQkptSHpIS0VGdz09
Password: 828w
Abstract
A common paradigm of learning robot policies is to rely on expert demonstrations. However, we often have limited access to expert demonstrations and collecting such data on robots with high degrees of freedom can be quite challenging. In practice, there are many other sources of human data that allow for learning robust robot policies or reward functions that go beyond learning from expert demonstrations.
I will talk about a set of techniques for learning from such non-traditional sources of data, i.e., play data, suboptimal demonstrations, low-dimensional action spaces, and language instructions. I will first discuss some of the challenges in the space of assistive robotics, and how we can benefit from more intuitive interfaces such as low-dimensional action spaces (latent actions) and language instructions in a shared autonomy paradigm. I will then talk about some of the challenges of moving from shared autonomy to full autonomy and specifically how to leverage play data or suboptimal demonstrations in such settings.
I will discuss PLATO, an algorithm that predicts latent affordances through object centric play, and then talk about how we can learn from suboptimal demonstrations by estimating the demonstrator’s expertise without access to any reward signals. I will finally wrap up the talk by demonstrating challenges of assistive robotics, specifically challenges in assistive feeding such as food acquisition and transfer and how they can benefit from the proposed approaches.
Relevant References
1) PLATO: Predicting Latent Affordances Through Object-Centric. Belkhale, Sadigh. CoRL 2022.
2) ILEED: Imitation Learning by Estimating Expertise of Demonstrators. Beliaev et al. ICML 2022.
3) LILA: Language-Informed Latent Actions. Karamcheti et al. CoRL 2021.
4) Controlling Assistive Robots with Learned Latent Actions. Losey et al. ICRA 2020.
Biography
Dorsa Sadigh is an assistant professor in Computer Science and Electrical Engineering at Stanford University. Her research interests lie in the intersection of robotics, learning, and control theory. Specifically, she is interested in developing algorithms for safe and adaptive human-robot and human-AI interaction. Dorsa received her doctoral degree in Electrical Engineering and Computer Sciences (EECS) from UC Berkeley in 2017, and received her bachelor’s degree in EECS from UC Berkeley in 2012. She is awarded the Sloan Fellowship, NSF CAREER, ONR Young Investigator Award, AFOSR Young Investigator Award, DARPA Young Faculty Award, Okawa Foundation Fellowship, MIT TR35, and the IEEE RAS Early Academic Career Award.
|
Event Information
