Safe Reinforcement Learning-Based Control of a Morphing Wing Aircraft

Welcome to the Institute of Flight Systems. Our work focuses on the interaction between aircraft configuration, pilots and modern flight system technology. From flight dynamics to unmanned aerial vehicles, from simulation to real flight tests - we analyse, test and develop innovations that will shape the flying of the future.

What to expect

Autonomous flight control increasingly relies on reinforcement learning (RL) to develop high-performance control strategies directly from data and interactions. RL can adapt online to new conditions and is particularly well suited for mission-level objectives that are difficult to design explicitly (efficiency, flight envelope, comfort).

In this context, a master’s thesis is offered to ensure the safe operation of an RL-based flight controller. The work includes the implementation of an actor-critic RL algorithm and a safety filter. Approaches from machine learning (including Gaussian processes) will be investigated to estimate uncertainties.
The thesis will be jointly supervised by DLR and Prof. Mayank Shekhar Jha from the Research Center for Automatic Control (CRAN), Centres Internationaux de Recherche (CNRS), Université de Lorraine, France.

Your tasks

• Design and implement a RL algorithm for attitude control of a morphing wing aircraft
• Train and evaluate the RL agent in simulation to achieve stable and efficient flight control performance
• Extend the RL based controller with a safety filter that prevents unsafe flight states such as excessive roll angles or aerodynamic stall
• Integrate uncertainity estimation within the safety filter framework, f.e. with Gaussian Processes
• Analyze and visualize training performance, safety margins and controller behavior
• Optional: Integrate the controller in the PROTEUS test aircraft and demonstrate in flight tests

Your profile

• Current enrollment in a Master’s program in Computer Science, Control Engineering or a related field
• Solid background in reinforcement learning, ideally policy gradient methods
• Good understanding of control engineering and flight dynamics
• Experience with TensorFlow, PyTorch for numerical optimization, neural networks and deep learning
• Strong programming skills, particularly in Python or C++

We look forward to getting to know you!

If you have any questions about this position (Vacancy-ID 3616) please contact:

Mark Spiller

Tel.: +49 531 295 1159