PS3Simulator: Physics-Parametrised Synthetic Sonar for Self-Supervised Sim-to-Real Transfer

Kamal Basha S, Athira Nambiar

Keywords: self-supervised learning, synthetic data generation, sim-to-real transfer, sonar image classification, vision transformers, masked image modeling, domain adaptation, underwater robotics, physics-based rendering

TL;DR: PS3Simulator generates physics-parametrised synthetic sonar data enabling structure-aware self-supervised learning for sim-to-real sonar classification without any real data collection.

Abstract:

Sonar imaging plays a critical role in underwater object detection and classification for maritime security, autonomous underwater vehicle operations, and environmental monitoring. However, real sonar datasets are scarce due to high collection costs, confidentiality restrictions, and expensive annotation, limiting the development of deep learning systems for underwater perception. Although synthetic data generation offers a scalable alternative, existing pipelines lack explicit physical parameters, and models trained on synthetic data suffer from a persistent sim-to-real domain gap. To address these limitations, three contributions are presented in this work. First, PS3Simulator,
A Physics parametrised synthetic Side-Scan Sonar dataset is black presented by incorporating per-image physical parameters including seabed material, grazing angle, altitude, and object rotation, enabling controllable dataset generation at scale. Second, the first application of I-JEPA~\cite{assran2023self} to the sonar domain proposes a structure-aware self-supervised learning framework based on masked latent prediction. Third, a cross-domain, cross-paradigm evaluation compares pretraining across ImageNet and PS3Simulator domains under a strict synthetic-train real-test protocol on KSLG and SCTD real-world sonar datasets. The proposed method achieves $70.9\%$ accuracy, outperforming DINO~cite{caron2021emerging} on identical PS3Simulator data by $+12.1\%$ ($\pm4.5\%$ vs $\pm11.9\%$ variance). At scale, I-JEPA ViT-H/14 attains $86.0\%$. These results confirm that structure-aware SSL transfers more effectively than appearance-based methods for sim-to-real sonar transfer without requiring real sonar data.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 4