The key results of this project will be:

• an innovative set of control-tutored learning strategies to orchestrate the behaviour of single and multi-agent systems
• their application to designing autonomous agents that can be deployed for SAR operations in natural disaster zones.

The project will advance knowledge with respect to the existing state of the art in several directions. First and foremost, it will combine control theoretic and machine learning approaches in an entirely novel way to achieve better control and learning performance, reducing the learning times and increasing data efficiency of the resulting algorithms. It will also contribute towards the development of safe and trusted machine learning algorithms, as, within the CT-MARL framework we propose, agents will inherit some of the convergence and robustness properties encoded in the control strategies used to inform the learning process. Also, the strategies will be contrasted with others existing in the literature, providing a set of testbed applications and validation results to be used as benchmarks well beyond the end of the project, to test future learning-based control solutions for multi-agent systems. 

Finally, the project will advance current know-how on designing autonomous multi-agent systems able to isolate and rescue people or animals in dangerous zones and corral them towards a safe zone in uncertain environments. This will be a step change with respect to existing solutions as the combination of control with machine learning will guarantee safe operation of the ensemble while endowing it with flexibility and the ability to adapt and self-organise, typical of natural multi-agent systems. The proposed application to the problem of steering schools of fish away from dangerous areas will also have an impact for environmental protection and mitigate possible natural disasters following antropic hazards such as oil spills in marine environments or to relocate animal groups following fires or floods.