High-Fidelity Deep-Sea Perception Using Simulation in the Loop

Abstract

Deep-sea robotic operations require a high level of safety, efficiency and reliability. In the development stage of such systems, measures have to be taken into account to validate performance in order to assess the achievement of these requirements. In the context of continuous system integration, we proposed a simulation-in-the-loop framework focusing on the mitigation of discrepancies between simulation and real-world conditions. While in our previous work we mainly targeted a high-fidelity simulation that embeds spatial conditions from recorded real-world data, this work emphasizes environmental conditions. We propose an optimization cycle which allows to enhance the fidelity of simulated underwater camera images in a backward optimization step and to enhance real-world images with knowledge available in simulation in a forward optimization step. Experimental results show that the proposed methodology optimized both simulation and real imagery, and subsequently ensures high fidelity.