Dynamical Distance Learning for Semi-Supervised and Unsupervised Skill Discovery

Explore the concept of Dynamical Distance Learning (DDL) in this informative video presentation. Learn how DDL functions as an auxiliary task for agents to determine distances between states in episodes, enhancing policy learning procedures. Delve into the paper's abstract, which outlines the challenges of manual reward function specification in reinforcement learning and introduces dynamical distances as a solution. Discover how this approach can be applied in semi-supervised settings, combining unsupervised environmental interaction with minimal preference supervision to achieve complex tasks. Examine the practical applications of DDL, including a real-world experiment involving a 9-DoF hand learning to turn a valve using raw image observations and only ten preference labels. Gain insights into how dynamical distances can provide well-shaped reward functions for new goals, potentially revolutionizing the efficiency of learning complex tasks in reinforcement learning.