Autonomous Unmanned Underwater Vehicles

• Autonomous UUVs are robotic systems that navigate underwater using onboard intelligence, robust control, and energy-efficient propulsion for diverse missions.
• They integrate advanced sensors, hybrid propulsion, and swarm coordination to achieve precise navigation, dynamic adaptation, and reliable communications in GNSS-denied environments.
• Ongoing research focuses on optimizing energy usage, adaptive control, and distributed coordination to overcome challenges in deep-sea operations and ensure fully autonomous performance.

Autonomous Unmanned Underwater Vehicles (UUVs) are robotic systems capable of executing missions in the underwater domain without direct human control, leveraging their own onboard intelligence, perception, and actuation. Autonomous UUVs include conventional Autonomous Underwater Vehicles (AUVs), gliders, bio-inspired bionic vehicles, underactuated multi-robot fleets, and recent hybrid architectures with collaborative or multi-agent capabilities. These platforms are fundamental for oceanography, environmental monitoring, defense, subsea inspection, and covert operations, where communication, navigation, and dynamic adaptation in the challenging underwater environment require advanced autonomy, robust control, and energy-efficient propulsion.

1. Classification, Architectures, and Physical Properties

Autonomous UUVs are classified by size, morphology, propulsion, and mission envelope. Size tiers (extracted from (Xu et al., 2024)):

Class	Mass / Payload	Depth Rating	Endurance	Example
Large	>500 kg / >10 kg	>500 m (up to 11,000 m)	Days	HUGIN Superior
Medium	≤500 kg / ≤10 kg	≤500 m (≤6000 m)	Hours–days	REMUS 6000
Small	<100 kg / <3 kg	<50 m	Hours	Xi’an Tianhe (micro-AUVs)

Core morphologies comprise torpedo-form (streamlined, high-speed), modular-body (swappable sensors, robotic arms), buoyancy-driven gliders (ultra-low-power, months-long missions), and hybrid AUV/UAVs (e.g., Nezha III, TJ-FlyingFish). Bionic platforms employing flapping fins enable improved maneuverability and bio-inspired operational modes (Zhou et al., 2024).

Energy systems are typically lithium-ion batteries (≈200 Wh/kg), with Peukert-law-limited discharge and onboard battery management/SoC estimation via extended Kalman filters. Buoyancy engines replace conventional thrusters in gliders, achieving vertical-to-horizontal transfer of motion at energy draw ≈1–5 W for long-range missions (Sollesnes et al., 2019). Deep-sea pressure-tolerant designs rely on oil-filled electronics and flexible bladders, eliminating rigid hulls to minimize weight and facilitate UAV retrieval/deployment (Sollesnes et al., 2019).

2. Hydrodynamics, Propulsion, and Efficiency

Hydrodynamic performance is governed by interactions between hull shape, local flow conditions, and propulsor-induced wakes. In rotating flow environments (e.g., group operations, shallow wave-current zones), drag coefficient $C_D$ increases nearly linearly with propeller RPM, up to four times above the baseline for a 0.5 m hull at $d=0.08$ m propeller distance. At 1200 RPM, $C_D \approx 0.085$ compared to $0.020$ (stopped), while spacing the propeller ≥$2D$ downstream or shrouding its wake reduces drag load (Mitra et al., 2021). Peak skin friction ($C_f$) and pressure coefficient ($C_p$) are observed at the propeller-facing end and decay toward the tail. Hull forms with gradual curvature and fairing designs mitigate sensitivity to swirl and incoming wake gradients.

Propulsion systems range from direct-drive brushless DC thrusters (10–200 W, high-efficiency, vectorable) to bionic flapping-fins whose cycle-by-cycle optimal gaits are determined by neural-network surrogates minimizing composite loss in thrust error, smoothness, and power (Zhou et al., 2024). A non-dimensional figure of merit (FOM), $\eta = F_{\rm avg} v_{\rm char}/P_{\rm avg}$, enables cross-design efficiency benchmarking. Flexible silicone (PDMS) fins outperformed rigid fins with up to 0.5 N thrust increase and 3 W per-cycle power reduction.

Energy efficiency improvements are pursued via adaptive thrust scheduling (mission-phase adapted), integration of fuel cells, and field-calibrated optimal parameter searches for inverse gait selection (Xu et al., 2024, Zhou et al., 2024).

3. Navigation, Localization, and Communication

Navigation without GNSS is intrinsic to stealth or deep-sea operation. In the absence of GPS, UUVs fuse Inertial Navigation System (INS), Doppler Velocity Log (DVL), and Long- or Ultra-Short Baseline (LBL/USBL) acoustic measurements using extended Kalman filtering (Xu et al., 2024). For covert operations, self-deployed beacon constellations (fixed or floating, deployed by UAV/USVs) enable ToA/TDoA multilateration, achieving sub-10 m accuracy over kilometer-scale missions with mean segment errors $<6.1$ m and overall 94% mission success in GNSS-denied environments (Albore et al., 22 Jan 2026).

Communication technologies are dominated by acoustic bands (1–10 kHz for tens of km, data rates 100 bps–20 kbps), with latency/throughput constrained by multipath, absorption, and bandwidth. Optical (UWOC) blue/green lasers yield Mbps rates but only up to 100 m and require clear water, while hybrid acoustic-optical schemes are under experimental validation (Xu et al., 2024).

In collaborative settings, surface–underwater interaction for high-accuracy positioning involves Fisher-information-optimized USV path planning (USBL) to maximize the Cramér–Rao bound for all AUVs’ positions (Xu et al., 2024). Distributed navigation architectures have demonstrated RMS positioning errors $>30$\% lower than fixed-base approaches, with robust performance sustained in simulated extreme sea conditions.

4. Motion Planning and Control Methods

The canonical 6-DOF nonlinear dynamic model, stacking Fossen’s notation, supports body-based velocities $(u,v,w,p,q,r)$ and pose $(x,y,z,\phi,\theta,\psi)$ (Chu et al., 2024, Zhu et al., 2022). Hydrodynamic torques and forces are partitioned into rigid-body, added-mass, Coriolis, and nonlinear damping components.

Global path planning employs grid-based (A*, D*, Dijkstra), visibility/graph, and sampling (PRM, RRT/RRT*) methods for obstacle avoidance, with local trajectory optimization via B-splines penalizing depth, thrust, turn, and collision violations (Zhu et al., 2022, MahmoudZadeh et al., 2020). Intelligent (GA, PSO, Ant Colony, BBO) and hybrid (APF+RL) algorithms enable scalable, multi-objective task- and route-assignment, with RL yielding energy savings by adaptive waypoint generation in unstructured terrains.

Control strategies include classical PID (rapid tuning, suboptimal for nonlinear/underactuated systems), Sliding Mode Control (robust, chattering-prone), Model Predictive Control (sys­tematic constraint handling, input/state adaptive, computationally intensive), and distributed optimal neurodynamics-based backstepping controllers for nonholonomic, underactuated formation consensus (Yan et al., 2023, Zhu et al., 2022). Backstepping with bio-inspired shunting models mitigates “explosion of terms” and improves practical robustness and boundedness in the presence of disturbances.

Autonomous multi-UUV assignment protocols encompass distributed consensus (min-consensus, formation-keeping), neural-network SOM for parallel assignment/path planning under current, and optimization-based (Hungarian/auction) approaches for linear assignments (Zhu et al., 2022). Differential-game-theoretic frameworks enable finite-time collaborative pursuit of maneuvering targets, with feedback Nash equilibria analytically characterized in the presence of communication delays and oceanic disturbances (Wei et al., 2022).

5. Multi-Agent Architectures, Swarms, and Planned Collaboration

Team autonomy and reactivity is achieved by hybrid reactive/deliberative architectures, such as ARMPA, which maintain three interconnected modules: (1) a deliberative mission planner (task/route optimization), (2) a reactive motion planner (real-time trajectory adaptation given mapped currents/obstacles), and (3) a synchronization module for execution-time correction and plan updating (MahmoudZadeh et al., 2020). This architecture is simulation-validated for missions up to 150 route nodes and real-time reactivity under dynamic conditions.

USV-AUV collaborative frameworks fuse Fisher information–driven surface navigation with distributed, reinforcement-learning–based AUV controllers (DDPG/SAC) for coordinated, robust, data-efficient localization and task execution—even under strong wave/vortex fields (Xu et al., 2024). Mission-level orchestration languages (e.g., Dolphin) introduce Groovy-hosted DSLs for compositional, event-driven tasking of UUV and UAV assets, supporting composable operators for sequential, parallel, conditional, and region-constrained execution (Lima et al., 2018).

Search-and-hunt scenarios adopt finite-time differential games, where UUV swarms optimize agent-level cost functions balancing collision-avoidance, consistency, and energy, in competition with evasive targets affected by environmental forcing and communication delays. Deep RL agents absorb unmodeled dynamics and delays, achieving practical convergence within several thousand episodes (Wei et al., 2022).

6. Challenges, Open Problems, and Future Directions

Autonomous UUV research must address several open challenges:

• Energy Efficiency: Battery and actuation limits necessitate improvements in ultra-low-power electronics, hybrid fuel systems, and vehicle-environment co-design, with glider-based and flexible-propulsion technologies as leading directions (Xu et al., 2024, Sollesnes et al., 2019, Zhou et al., 2024).
• Robust Underwater Localization: Deep-water, GNSS-denied environments call for hybrid acoustic–optical systems, cooperative beacon networks, and sensor-fusion with distributed Kalman filtering. Infrastructure-free, attitude-enhanced state estimation for single or swarming UUVs remains crucial (Albore et al., 22 Jan 2026, Liu et al., 16 Jun 2025).
• Adaptation to Environmental Variability: Onboard autonomy must dynamically reconfigure control gains and mission plans under changing current, temperature, and biofouling conditions. Formal software verification (sense-plan-act pipelines) is required to guarantee safety in long-duration unsupervised operations (Xu et al., 2024).
• Fully Autonomous Operations and Docking: Certified autonomy stacks and robust passive/active docking mechanisms for energy and data transfers are under development (Xu et al., 2024).
• Distributed Coordination and Learning: Multi-agent UUV fleets need scalable consensus/formation algorithms resilient to communication constraints and delays, supported by theoretical bounds on convergence and performance (Yan et al., 2023, Wei et al., 2022).
• Human-Machine Teaming: Research is moving towards reference frameworks where autonomy is measured and allocated at the function level, enabling seamless operator oversight and exception management (Atyabi et al., 2020).

Hybrid architectures that combine simulation-based RL pretraining (fast, domain-randomized), model-based control, and residual learning are maturing to bridge the sim-to-real gap (Chu et al., 2024). Open simulation environments supporting multi-agent RL, sensor models, and wide-environmental parameterization are extending the design and autonomy landscape for next-generation UUVs.

7. Representative Applications and Performance Metrics

Autonomous UUVs support applications in oceanography, infrastructure inspection, habitat mapping, defense/reconnaissance, and disaster response. Modular micro-AUVs (e.g., SeaShark) deliver <5% navigation drift over 100 m, centimeter-scale imaging for photogrammetry, and stackable/rotatable payloads for multi-modal sensing with passive buoyancy (Christensen et al., 2020).

Precision in stealth navigation via self-deployed beacons achieves mean segment errors ≈4.8–6.1 m, <10 m mission drift, energy usage ≈8 kWh/run, while enforcing average surfacing rates <2. Formation-tracking fleets under distributed/optimal control maintain sub-0.1 m RMS error under nonholonomic/underactuated constraints (Liu et al., 16 Jun 2025, Yan et al., 2023). RL-driven tracking control using high-fidelity physics simulation allows trajectory error convergence under 0.02 m and wall-clock training in minutes for classic, helical, and lemniscate paths (Chu et al., 2024).

Flexible-finned UUVs, optimized with deep learning surrogate models and constraint-based inverse search, achieve thrust and power tracking improvements of up to 0.5 N/3.0 W per movement cycle, translating to 30–50% increases in practical endurance (Zhou et al., 2024). Swarm hunting controllers maintain high mission success rates and convergence to Nash equilibria despite disturbance and delay, showing the resilience required for real-time multi-agent operations (Wei et al., 2022).