Learning with Humans in the Loop

Gokul Swamy

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2020-76
May 28, 2020

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-76.pdf

Robots are starting to leave factory floors and enter daily life, sharing our roads and homes. For this transition to be a smooth one, robots need to be able to learn to interact safely with the highly varied environments and other agents that compose the real world. Much of this task, at some level, requires learning from people, either by (1) developing predictive models of human behavior to be used in motion planning, (2) imitating expert demonstrations produced by people, or (3) learning from user feedback, often given as labels for data. With humans in the loop, learning algorithms should be able to learn quickly from data that is both easy and safe for people to provide. We present methods and experimental evidence that are a useful foundation for developing scalable algorithms for each of these tasks. In Chapter 2, we present experimental evidence for the advantages of a theory-of-mind inductive bias for human motion prediction in the autonomous vehicle domain. In Chapter 3, we propose an algorithm for scaling up assistive teleoperation to multiple robots and validate its efficacy with simulated experiments, a user study, and a hardware demonstration. In Chapter 4, we describe a proof-of-concept system for brain-computer-interface-based shared autonomy via deep learning and present preliminary simulated experiments. In Chapter 5, we provide some conclusions and situate our work in a larger social and ethical context.

Advisor: Anca Dragan


BibTeX citation:

@mastersthesis{Swamy:EECS-2020-76,
    Author = {Swamy, Gokul},
    Editor = {Dragan, Anca and Levine, Sergey},
    Title = {Learning with Humans in the Loop},
    School = {EECS Department, University of California, Berkeley},
    Year = {2020},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-76.html},
    Number = {UCB/EECS-2020-76},
    Abstract = {Robots are starting to leave factory floors and enter daily life, sharing our roads and homes. For this transition to be a smooth one, robots need to be able to learn to interact safely with the highly varied environments and other agents that compose the real world. Much of this task, at some level, requires learning from people, either by (1) developing predictive models of human behavior to be used in motion planning, (2) imitating expert demonstrations produced by people, or (3) learning from user feedback, often given as labels for data. With humans in the loop, learning algorithms should be able to learn quickly from data that is both easy and safe for people to provide. We present methods and experimental evidence that are a useful foundation for developing scalable algorithms for each of these tasks. In Chapter 2, we present experimental evidence for the advantages of a theory-of-mind inductive bias for human motion prediction in the autonomous vehicle domain. In Chapter 3, we propose an algorithm for scaling up assistive teleoperation to multiple robots and validate its efficacy with simulated experiments, a user study, and a hardware demonstration. In Chapter 4, we describe a proof-of-concept system for brain-computer-interface-based shared autonomy via deep learning and present preliminary simulated experiments. In Chapter 5, we provide some conclusions and situate our work in a larger social and ethical context.}
}

EndNote citation:

%0 Thesis
%A Swamy, Gokul
%E Dragan, Anca
%E Levine, Sergey
%T Learning with Humans in the Loop
%I EECS Department, University of California, Berkeley
%D 2020
%8 May 28
%@ UCB/EECS-2020-76
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-76.html
%F Swamy:EECS-2020-76