This document summarizes a research paper about using reinforcement learning to extract robotic skills from human feedback. The key points are: 1) Previous research on reward design for reinforcement learning requires significant engineering costs, while human-in-the-loop RL allows acquiring skills through interactive feedback during training to remove hand-engineered rewards. 2) This paper aims to address the issue that complex tasks require unrealistically large amounts of interactive feedback to acquire good policies. 3) The proposed method trains a preference classifier on a small subset of labeled data to learn human preferences, then trains an encoder-decoder to extract skills from trajectories while following the learned human preferences.