Beyond Shallow-Water Photorealism: Physically and Sensor-Grounded Simulation for Deep-Sea Robotics
本文针对深海环境下物理和传感器真实感不足的问题,通过扩展Stonefish模拟器,引入多种物理和环境因素,提高深海机器人仿真准确性。
本文针对深海环境下物理和传感器真实感不足的问题,通过扩展Stonefish模拟器,引入多种物理和环境因素,提高深海机器人仿真准确性。
为解决水下采样时自主水下航行器状态估计不准的问题,提出了一种基于接触辅助因子图的定位框架,融合了吸盘接触事件、自适应视觉里程计及传感器信息。
This study addresses the challenge of efficiently quantifying uncertainty in three-dimensional seismic traveltime tomography, which is hindered by ill-posed inversion and the curse of dimensionality. The authors propose a mesh-free Bayesian approach that integrates physics-informed neural networks with neural field representations of velocity structure, enabling efficient posterior sampling via functional-space particle variational inference. Analytical marginalization over passive-source parameters is incorporated to enhance computational tractability. This framework represents the first scalable Bayesian tomography method capable of jointly handling large-scale active- and passive-source data, substantially improving scalability and data efficiency. Validations on synthetic and real-world datasets from the Kii Peninsula offshore region successfully recover key geological structures with well-calibrated uncertainties; posterior source locations exhibit vertical offsets of 10–15 km, consistent with prior studies, while drastically reducing model storage requirements.
This study addresses the challenge of distinguishing abrupt seagrass loss from seasonal fluctuations by integrating aerial photographs since the 1940s, high-resolution satellite imagery, GRUS data, and Sentinel-2 monthly composites to reconstruct an 80-year history of seagrass coverage in Ago Bay, Japan—the first such effort combining multi-source remote sensing with deep learning. The authors propose a novel monitoring paradigm incorporating seasonal normalization and extreme anomaly detection, employing a YOLO-based segmentation model and time series analysis to achieve reconstruction accuracy exceeding 0.9. The analysis confirms that the dramatic decline to 0.2 hectares in 2025 constitutes an anomalous event, most likely driven by anomalously high summer sea surface temperatures, thereby refining Essential Ocean Variables (EOVs) and reference conditions for seagrass ecosystem assessment.
This work addresses the challenges of ambiguous coupling between physical structure and learning models, energy non-conservation, and poor extrapolation in mesh-based modeling of continuous physical systems. The authors propose Mesh Field Theory and its neural implementation, MeshFT-Net, which uniquely decouples topological and metric structures within mesh representations. By formulating dynamics in port-Hamiltonian form and learning only the metric-dependent component—while enforcing conservative interconnections dictated solely by topology—the approach embeds strong physically consistent inductive biases. The network architecture integrates principles from port-Hamiltonian systems, locality, permutation equivariance, directional covariance, and energy dissipation constraints. Experiments demonstrate that the method achieves near-zero energy drift, accurate dispersion relations, and exact momentum conservation across diverse physical systems, while exhibiting superior extrapolation capability and high data efficiency.
本文针对深海环境下物理和传感器真实感不足的问题,通过扩展Stonefish模拟器,引入多种物理和环境因素,提高深海机器人仿真准确性。
为解决水下采样时自主水下航行器状态估计不准的问题,提出了一种基于接触辅助因子图的定位框架,融合了吸盘接触事件、自适应视觉里程计及传感器信息。
This study addresses the challenge of efficiently quantifying uncertainty in three-dimensional seismic traveltime tomography, which is hindered by ill-posed inversion and the curse of dimensionality. The authors propose a mesh-free Bayesian approach that integrates physics-informed neural networks with neural field representations of velocity structure, enabling efficient posterior sampling via functional-space particle variational inference. Analytical marginalization over passive-source parameters is incorporated to enhance computational tractability. This framework represents the first scalable Bayesian tomography method capable of jointly handling large-scale active- and passive-source data, substantially improving scalability and data efficiency. Validations on synthetic and real-world datasets from the Kii Peninsula offshore region successfully recover key geological structures with well-calibrated uncertainties; posterior source locations exhibit vertical offsets of 10–15 km, consistent with prior studies, while drastically reducing model storage requirements.
This study addresses the challenge of distinguishing abrupt seagrass loss from seasonal fluctuations by integrating aerial photographs since the 1940s, high-resolution satellite imagery, GRUS data, and Sentinel-2 monthly composites to reconstruct an 80-year history of seagrass coverage in Ago Bay, Japan—the first such effort combining multi-source remote sensing with deep learning. The authors propose a novel monitoring paradigm incorporating seasonal normalization and extreme anomaly detection, employing a YOLO-based segmentation model and time series analysis to achieve reconstruction accuracy exceeding 0.9. The analysis confirms that the dramatic decline to 0.2 hectares in 2025 constitutes an anomalous event, most likely driven by anomalously high summer sea surface temperatures, thereby refining Essential Ocean Variables (EOVs) and reference conditions for seagrass ecosystem assessment.
This work addresses the challenges of ambiguous coupling between physical structure and learning models, energy non-conservation, and poor extrapolation in mesh-based modeling of continuous physical systems. The authors propose Mesh Field Theory and its neural implementation, MeshFT-Net, which uniquely decouples topological and metric structures within mesh representations. By formulating dynamics in port-Hamiltonian form and learning only the metric-dependent component—while enforcing conservative interconnections dictated solely by topology—the approach embeds strong physically consistent inductive biases. The network architecture integrates principles from port-Hamiltonian systems, locality, permutation equivariance, directional covariance, and energy dissipation constraints. Experiments demonstrate that the method achieves near-zero energy drift, accurate dispersion relations, and exact momentum conservation across diverse physical systems, while exhibiting superior extrapolation capability and high data efficiency.