Master's Thesis: Improving Safety and Efficiency of Off-Policy RL Using Human Interventions

TL;DR

Investigated how human interventions can improve the safety and sample efficiency of off-policy reinforcement learning for robotics, receiving a perfect grade.

The thesis studies human-in-the-loop methods for off-policy reinforcement learning, focusing on safer and more sample-efficient robot learning.

The work evaluates learning from human interventions experimentally and forms the basis for continued research toward a scientific publication at FZI.