Data-efficient deep reinforcement learning for robotics
Date and time: 13 January 2022, 12:00 – 13:00 CET (UTC +1)
Speaker: Ali Ghadirzadeh, RPL & Stanford
Title: Data-efficient deep reinforcement learning for robotics
Meeting ID: 695 6088 7455
Host: Puzhao Zhang
Watch the recorded presentation:
Abstract: Reinforcement learning (RL) is the study of learning action-selection policies through interactions and trial and error. Deep RL refers to algorithms that combine RL policy training with the modelling capabilities of deep learning methods. In recent years, deep RL has achieved superhuman performance in solving very complex decision-making problems such as the game of Go and Dota2. However, such solutions do not scale well to robotics because of the cost of real-robot samples, safety concerns of real-robot random exploration, and the diversity of tasks in open-ended robotic applications.
This talk presents main challenges in robot learning research and reviews some of the recent advancements towards making RL algorithms more well-suited to robotics by providing experimental results on learning/exploiting (world) models, training generative models to make the best use of prior knowledge, and learning common structures between different robotic tasks. Future research directions and open problems in robotics research will conclude the talk.