Simulation-based Reinforcement Learning (SiRL)

Combining physical simulation with machine learning opens up the possibility of solving a large repertoire of challenges that previously have been impossible to solve. It is a future-enabling technology. This simulation-based learning approach is highly relevant in scenarios where data is limited or not easily available. In such a scenario the simulator provides a model or prior for the learning process. The learning-based process can on the other hand bridge the gap to reality as it allows the simulator to incorporate knowledge of the real world better through data. In combination, both flavours combined can provide solutions for the control of very advanced systems, typical robots, for solving quite complex problems. As a whole, this field of simulation-based learning addresses many interesting issues such as motion control, motion imitation, and system identification like parameter estimation, or digital twinning, for instance creating data-driven models of humans, etc. These are highly relevant for solving actual real-world challenges. Such as advanced autonomous robotics manipulation and navigation, prediction or detection of human motion for improved health-care services, creation of realistic computer animations for XR, movies, and computer games, or in digital twining such as learning the physics of the carbon dioxide cycle in soil which is an important aspect of predicting climate consequences from farming. Many more examples exist. This course will prepare students for working with this world-changing technology.

This course is ideal for students who wish to work with simulation-based learning in their master's thesis. It is particularly ideal for students who want to work with robotics or character animation, or content creation for XR.

Students are expected to solve weekly practical exercises mostly programming with python, but some C++ could be involved, as well as some lab-space work as well. Students are expected to be self-driven in getting the software to work on their own devices, like installing an auto differentiation framework and making it run. Students could be exposed to working with vision or motion capture data.