Public defence in Automation, Systems and Control Engineering, M.Sc. Karol Arndt

M.Sc. Karol Arndt will defend the thesis "Safe and efficient transfer of robot policies from simulation to the real world" on 16 May 2023 at 12 (EET) in Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation, in lecture hall T2, Konemiehentie 2, Espoo.

Opponent: Prof. Markus Vincze, TU Wien, Austria
Custos: Prof. Ville Kyrki, Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation

Over the past decade, there has been great progress in training intelligent agents to perform various tasks through reinforcement. However, while achieving good results in virtual environments, learning methods still underperform in real-world scenarios where collecting data used for training is costly and difficult. To address this, it has been proposed to train agents in simulation and to later transfer the learned skills to the real world. There are, however, various challenges that need to be addressed to perform the skill transfer in a safe and data-efficient manner. 

To achieve this, the thesis proposes and discusses a range of methods for training robots using simulation and deploying the learned behavior to the real robot. The proposed methods can help robots learn faster and avoid safety risks related to training in the real world. 

The thesis formalizes this problem and breaks it down into smaller pieces. It then explores different ways to address each piece, such as how to transfer behavior policies from simulation to the real world without any real-world data, or how to use real-world data to improve the accuracy of the transfer. The work also addresses important safety concerns that arise when transferring learned behaviors to the real world, such as making sure the robot does not take actions that could cause harm. It proposes a way to address these concerns through safety-aware domain transfer algorithms. 

The research shows that real robots can be successfully trained using simulations, and that the learned behavior can be transferred to the real world with good accuracy in a safe and efficient manner, improving upon the current ways robots can be trained to perform complex tasks in the real world.

Thesis available for public display 10 days prior to the defence at: https://aaltodoc.aalto.fi/doc_public/eonly/riiputus/
Doctoral theses in the School of Electrical Engineering: https://aaltodoc.aalto.fi/handle/123456789/53