System-Level Optimization of AUV Swarm Control and Perception: An Energy-Aware Federated Meta-Transfer Learning Framework with Digital Twin Validation

Zinan Nie, Hongjun Tian, Yijie Yin, Yuhan Zhou, Wei Li, Yang Xiong, Yuyi Wang, Z. Zhang, Xinyu Du, Dongxiao Xie, Manlin Wang, Shijie HuangJournal of Marine Science and Engineering 2026 · DOI

TL;DR

This paper introduces a novel Federated Meta-Transfer Learning (FMTL) framework for optimizing AUV swarms in deep-sea exploration. It tackles the challenges of limited communication, energy constraints, and sensor heterogeneity by tightly integrating perception, explainability, and adaptive control, achieving superior performance and efficiency.

What Problem Does This Paper Address?

Deep-sea exploration using Autonomous Underwater Vehicles (AUVs) faces significant hurdles. Deployments are constrained by limited energy and unreliable, low-bandwidth acoustic communication, making mission-level coordination difficult. Current systems often treat perception and control as separate modules, leading to inefficient sensing, communication, and degraded performance, especially when dealing with diverse sensors and changing geological conditions. This paper aims to bridge the gap between isolated data acquisition and collaborative, in-situ knowledge formation, enabling AUV swarms to adapt their exploration strategies in real-time.

Methods

The core of this work is the Federated Meta-Transfer Learning (FMTL) framework, which operates hierarchically:

  1. Representation Learning: A mapper-centric aggregation strategy is employed for federated learning. Unlike standard approaches that average all model parameters, FMTL keeps sensor-specific encoders local and only aggregates a shared "mapper" module. This aligns heterogeneous sensor data into a unified latent space using a physics-informed contrastive objective, significantly reducing communication overhead and enabling cross-modality fusion without requiring raw data sharing.
  2. Meta-Learning Adaptation: This stage enables rapid adaptation to new environments with minimal data. It uses a context encoder to infer a latent description of the local environment, conditioning the model's inference without extensive retraining.
  3. Energy-Aware Control: This integrates Federated Explainable AI (FXAI) for trustworthy reasoning and decentralized Multi-Agent Reinforcement Learning (MARL) for adaptive path planning under energy constraints. FXAI provides attribution maps to understand model decisions, while MARL optimizes exploration strategies.

Datasets

  • Pre-training: Publicly available global marine geophysical data, including GEBCO 2024 Grid (bathymetry), EMAG2v3 (magnetic anomaly), and hydrothermal vent field data (SMS).
  • Validation:
    • A digital-twin seafloor environment simulating geological heterogeneity, sensor diversity, and acoustic communication limitations.
  • Hardware-in-the-Loop (HIL) validation platform using NVIDIA Jetson AGX Orin modules and S2CR 18/34 acoustic modem emulators.
  • Real-world geophysical datasets for supplementary validation.

Key Insights and Ideas

  • Mapper-Centric Aggregation is Crucial for Heterogeneity: Standard federated learning struggles with sensor heterogeneity because averaging encoder parameters leads to "semantic mismatch." FMTL's approach of keeping encoders local and aggregating a shared mapper is key to effective cross-modality fusion.
  • Interpretability as a Trust Mechanism: FXAI provides a "visual consensus" of why the swarm believes a target exists, transforming the swarm from a data collector into a collaborative partner. This is vital for operational trust and reducing costly sampling actions.
  • Meta-Learning for Rapid Adaptation: In the dynamic deep-sea environment, meta-learning is essential for quickly adapting to new geological regimes without extensive retraining, significantly improving "time-to-effectiveness."
  • Energy-Intelligence Trade-off is Paramount: For long-endurance AUV missions, optimizing for "Information Gain per Joule" is as critical as accuracy. FMTL demonstrates a significant improvement in this metric.
  • Tiered Deployment Protocol: A practical operational guideline is proposed, using a "loosely coupled" mode for initial exploration and switching to the full FMTL protocol for target verification only when confidence and multi-modal evidence are high.

Key Results

  • Superior Target Identification: FMTL achieved a dominant AUC of 0.94 for target identification, significantly outperforming baselines like FedDRL-AUV (0.86) and GNN-Mapper (0.88).
  • Communication Efficiency: FMTL achieved 90% peak performance with 500 MB of communication, a substantial reduction compared to centralized approaches. An 8-bit quantization strategy further reduced communication cost to 62.5 MB while retaining 98% of baseline accuracy.
  • Rapid Adaptation: Meta-learning enabled FMTL to reach 90% performance in 5 missions in a new geological province, an 8.4x improvement over standard fine-tuning.
  • Energy Efficiency: FMTL demonstrated a 4.5x improvement in information gain per Joule compared to centralized streaming and 1.8x improvement over standard FedAvg.
  • Robustness: The framework showed resilience to sensor failures, intermittent corruption, and realistic acoustic channel impairments, with performance degrading gracefully rather than catastrophically.
  • FXAI Impact: FXAI-Gated policies reduced false-positive sampling actions by 22%, leading to a 14% saving in total mission energy consumption.

Limitations and Future Work

  • Operational Deployment: The current work is a proof-of-concept. Field trials are required to validate end-to-end feasibility under real sensing, communication, and coordination coupling.
  • Physical Hydrodynamics and Multipath Effects: Open-ocean trials are needed to verify behavior under these conditions.
  • Delay-Tolerant Planning: Future work will focus on developing planning with explicit feasibility and reachability guarantees.
  • Asynchronous Consensus: Further research is needed for asynchronous consensus-based formation maintenance.
  • Extreme Information Sparsity: The "Sparse-Deceptive" environments revealed a "posterior collapse" issue in meta-learning. Future work should incorporate active, curiosity-driven exploration to better handle such scenarios.
  • Long-Term Drift and Reliability: While drift correction was demonstrated, long-term sensor drift and reliability under prolonged communication loss remain open issues.

Summary updated

Abstract

Deep-sea exploration increasingly relies on Autonomous Underwater Vehicles (AUVs) to enable persistent, wide-area surveying in harsh and uncertain environments. In practice, however, deployments are constrained by tight energy budgets and bandwidth-limited, intermittent acoustic links, which complicate mission-level coordination. Moreover, many existing systems treat perception and control as loosely coupled modules, often resulting in redundant sensing, inefficient communication, and degraded overall performance—particularly under heterogeneous sensing modalities and shifting geological conditions. To address these challenges, we propose a hierarchical Federated Meta-Transfer Learning (FMTL) framework that tightly integrates collaborative perception with adaptive control for swarm optimization. The framework operates at three levels: (1) Representation Learning aligns heterogeneous sensors in a shared latent space via a physics-informed contrastive objective, substantially reducing communication overhead; (2) Meta-Learning Adaptation enables rapid transfer and convergence in new environments with minimal data exchange; and (3) Energy-Aware Control realizes closed-loop exploration by coupling Federated Explainable AI (FXAI) with decentralized multi-agent reinforcement learning (MARL) for path planning under energy constraints. Validated in high-fidelity hardware-in-the-loop simulations and a digital-twin environment, FMTL outperforms state-of-the-art baselines, achieving an AUC of 0.94 for target identification. Furthermore, an energy–intelligence Pareto analysis demonstrates a 4.5× improvement in information gain per Joule. Overall, this work provides a physically consistent and communication-efficient blueprint for the optimization and control of next-generation intelligent marine swarms.
BibTeX
@article{nie2026system,
  author = {Zinan Nie and Hongjun Tian and Yijie Yin and Yuhan Zhou and Wei Li and Yang Xiong and Yuyi Wang and Z. Zhang and Xinyu Du and Dongxiao Xie and Manlin Wang and Shijie Huang},
  title = {{System-Level Optimization of AUV Swarm Control and Perception: An Energy-Aware Federated Meta-Transfer Learning Framework with Digital Twin Validation}},
  journal = {Journal of Marine Science and Engineering},
  year = {2026},
  doi = {10.3390/jmse14040384},
}

Loading related content...

Figure