Towards Reinforcement Learning Control of an Aerial Robot and Physical Interaction Tasks

Aerial robots are increasingly used for applications such as infrastructure inspection, environmental monitoring, and disaster response. While most systems operate in free flight, many emerging tasks require aerial robots to physically interact with their environment. Such interaction introduces significant challenges, as contact forces can destabilize the aerial platform and require precise control of motion and force. This thesis investigates the use of reinforcement learning (RL) for controlling a fully actuated aerial robot with the long-term goal of enabling aerial physical interaction tasks. A reinforcement learning framework is developed in which the control policy is trained in simulation using a progressive learning strategy. The training process begins with velocity stabilization and gradually extends to altitude control, position tracking, and stable hovering. The proposed approach is implemented using physics-based simulation environments that model the robot dynamics, actuator behavior, and contact interactions. The results demonstrate that reinforcement learning can successfully learn stable hovering control for a fully actuated aerial platform. These results establish the essential components required for future research on aerial physical interaction. Errata: Time axes in select figures were inadvertently labelled in simulation timesteps rather than seconds; the intended unit is seconds throughout, and no results or conclusions are affected by this typographical error.