Learning Robot Policies from Imperfect Teachers - Taylor Kessler Faulkner, UT Austin

Explore a lecture on developing robot learning algorithms that can effectively utilize input from imperfect human teachers. Delve into the challenges of interactive Reinforcement Learning when dealing with inattentive or inaccurate human instructors. Discover innovative approaches that enable robots to learn with or without constant human attention and to make use of both correct and incorrect feedback. Gain insights into Boltzmann Exploration, Policy Shaping, and trust calculation methods. Examine the results of human studies and simulations that demonstrate the effectiveness of these algorithms. Consider the implications for making robot teaching more accessible to non-experts and the potential for future research in this field.

Introduction
Welcome
Learning from Humans
Boltzmann Exploration
Policy Shaping
Human Study
Prior Algorithms
Bad Feedback
Prior Knowledge
Trust
Demonstrations
Repair
Repair Example
Trust Calculation
Comparison
Simulation
Feedback Quality
Feedback Reliability
Example
Experiment
Summary
Active Amps
Future Research
Current Work
Audience Questions
Thanks