From Camera Image to Active Target Tracking: Modelling, Encoding and Metrical Analysis for Unmanned Underwater Vehicles

Traditional marine mammal `tagging' involves attaching sensors such as GPS to the bodies of the former. A less intrusive approach exploits \gls{UUV}. The growth and development of AI in autonomous underwater vehicle navigation enables the modelling of training environments in simulation. Game engines, an increasingly useful tool, provide visual and physical fidelity suitable for sim-to-real applications. Other solutions including \gls{UVMS} and \gls{L2D} exploit such simulations for sim-to-real transfer, providing only satisfactory results, where poor environment generalisation leads to incorrect command signals. We exploit a state-of-the-art \gls{CMVAE} enhancing such a generalisation. Few solutions (if any) provide the desired results for underwater active target tracking. We propose \gls{SWIM2}, coupling a Unity simulation modelling a \gls{BLUEROV} underwater rover with a \gls{DRL} backend. Common evaluation of \gls{RL} algorithms analyses reward coupled with naked eye visual analysis, not completely capturing the desired behaviour. We introduce custom behaviour metrics linking our objectives and unprecedented in current ROV simulators, assessing rover behaviour and providing meaningful comparisons against competitors. To reduce the load on power and hardware resources, \gls{SWIM2} exploits image data alone from a modelled \gls{BLUEROV} camera. A state-of-the-art \gls{CMVAE} enables a dimensionality reduction while minimising information loss. Our experiments show that SAC maximises our objectives.