Robin Menzenbach, "Benchmarking Sim-2-Real Algorithms on Real-World Platforms", Technical University of Darmstadt, 2019.

Learning from simulation is particularly useful, because it is typically cheaper and safer than learning on real-world systems. Nevertheless, the transfer of learned behavior from the simulation to the real word can impose difficulties because of the so-called ’reality gap’. There are multiple approaches trying to close the gap. Although many benchmarks of reinforcement learning algorithms exist, state-of-the-art sim-2-real methods are rarely compared. In this thesis, we compare two recent methods on Furuta pendulum swing up and ball balancing tasks. The performed benchmarks aim at assessing sim-2-sim and sim-2-real transferability. We show that the application of sim-2-real methods significantly improves the transferability of learned behavior.