Efficient Robot Skill Learning via Grounded Simulation Learning, Imitation Learning - Peter Stone

Explore efficient robot skill learning techniques in this comprehensive seminar on theoretical machine learning. Delve into grounded simulation learning, imitation learning from observation, and off-policy reinforcement learning as presented by Peter Stone from The University of Texas at Austin. Discover the evolution of RoboCup Soccer and RoboCup@Home, and examine the challenges of applying reinforcement learning to physical robots. Investigate the concept of grounded simulation learning, including simulator grounding and grounded action transformation. Learn about importance sampling policy evaluation, behavior policy search problems, and the optimal behavior policy. Gain insights into regression importance sampling and the process of transferring robot skills from the real world to simulations and back.

Intro
Research Question
Efficient Robot Skill Learning
RoboCup Soccer
RoboCup 1997-1998
RoboCup@Home
Reinforcement Learning for Physical Robots
Reinforcement Learning in Simulation
Grounded Simulation Learning
Simulator Grounding
Grounded Action Transformation
Supervised Implementation
GSL Summary
Importance Sampling Policy Evaluation
RL Importance Sampling Myths
Behavior Policy Search Problem
The Optimal Behavior Policy
Behavior Policy Gradient Theorem
Regression Importance Sampling
Robot Skill Learning: Real World to Sim and Back