Safety in Transformer-based Reinforcement Learning

MSc assignment

Integration of machine learning in robotics is demonstrating tremendous impact in dealing with complex and uncertain tasks that cannot be tackled with classical model-based control techniques. The use of machine learning schemes for robots acting in unpredictable environments must then come together with guarantees of prevention of critical hazards along the execution of the tasks.

Recently Transformers, a class of highly expressive deep learning models primarily developed for processing sequential data, has achieved great success in tasks involving Natural Language Processing, Computer Vision, and more recently, in sequential decision-making in Reinforcement Learning (RL). Their strength lies in handling long-range dependencies within data, allowing them to make sense of sequences, whether they are sentences, series of images, or even decision-making sequences in a control task.

The research aims to explore the possibility of using Transformers in the Reinforcement Learning framework, with the goal of exploring architectures producing both expressive models and strong safety guarantees in robotics applications.