Current Research Activity

First pillar: My work is about guiding a Bayesian Optimizer with information sources of limited accuracy. These low-fidelity information sources include first-principle based models that capture the overall behavior of a complex target system. The objective is to investigate how far we can benefit from these available sources to improve the data efficiency of iterative learning controllers applied to a real-world system. Such systems include complex manufacturing setups where acquiring real measurement data is costly. My methodology aims to leverage low-fidelity digital twins as additional data sources to alleviate the cost of tuning the system.

Second pillar: I train a model-free reinforcement learning agent to learn the mismatch between a simulation and the real dynamics of a robot arm. We built a physical robot cell where a multi-degree-of-freedom arm tracks a moving object using online camera feedback. The central question is: can we learn to model the uncertainty in simulation well enough to successfully transfer the agent to the real robot—thus compensating for simulation inaccuracies?