Theses

In reinforcement learning (RL), an agent learns to optimally interact with its environment through a trial-and-error approach. There are numerous success stories of RL, for instance in the games of Chess or Go, but also various video games, where superhuman performance coulde be achieved. In the context of technical systems such as robots or autonomous vehicles, however, there are additional challenges, since it is not possible to perform arbitrarily many experiments on the real system, in particular if safety cannot be guaranteed. A common approach is thus to instead train on a simulation model, and then transfer the learned policy to the real system. However, there is a sim-to-real gap. This means that the model is never 100% accurate such that the learned policy can be suboptimal or even infeasible for the real system.

The task of this thesis is to study, implement and validate the GARAT algorithm (Generative Adversarial Reinforced Action Transformation, https://proceedings.neurips.cc/paper_files/paper/2020/file/28f248e9279ac845995c4e9f8af35c2b-Paper.pdf), which aims at overcoming the sim-to-real gap by modifying the action such that it successfully works on the real system, while requiring only a very small number of interactions with the real system.

Extension of a symmetric reinforcemet learning framework (known as homomorphic MDPs) from discrete symmetries (flips, 90° rotations) to continuous ones, with applications in robotics or continuum/fluid mechanics.