This work focuses on learning and experimentally testing of safe physical action control policies on underactuated aerial robots through Deep Reinforcement Learning (DRL) methods. These DRL policies are trained
with model-free methods through the deployment of Proximal Policy Optimisation (PPO) within a Markov Decision Process (MDP) framework. Additional safety features are embedded into the policies which allow for
performing physical interaction tasks safely, even while losing physical contact with the manipulated object.
Experimental testing is conducted in an indoor drone-lab test environment equipped, where tracking is achieved through the Opti-track system with sub-mm precision in position.