Betreuer:
Heiko Donat

Transfer Learning is a current frontier in machine learning, enabling the learning of complex patterns from data by reusing already learned mappings, significantly reducing the amount of data to learn a good approximation, and enabling efficient approaches like one shot learning.
Therefore, transfer learning is especially interested in the realms of robotics as data is scarce, because producing it with a real robot may damage the robot by wear and tear or taking too much time to be a realistic option.
Due to the advances in model-based physics-simulation data can be produced in large in simulation, while transfer learning enables the adaption to real robots.
In this seminar thesis, an overview about the current hot topics in transfer learning for robot applications should be given. The main references are listed below, but may be extended by the presenter. The seminar thesis and talk should have the focus on which type of transfer learning yields the best application in the robotics domain.

[1] https://www.mdpi.com/2079-9292/10/12/1491
[2] https://ruder.io/transfer-learning/
[3] https://arxiv.org/pdf/1808.01974.pdf
[4] https://tspace.library.utoronto.ca/bitstream/1807/70527/1/Raimalwala_Kai...