Autonomous Underwater Cognitive Systems

• AUCS is a cognitive system that continuously senses, understands, plans, acts, and learns to support complex underwater missions.
• It employs a layered architecture combining reactive control, deliberative planning, and online learning, as demonstrated in frameworks like DINOS-R and UROSA.
• AUCS addresses challenges in energy constraints, bandwidth limitations, and dynamic environmental conditions through advanced decision-making and sensor fusion.

Autonomous Underwater Cognitive System (AUCS) denotes an integrated underwater autonomy paradigm that closes a sense–understand–plan–act–learn loop under the physical, communicative, and energetic constraints of subsea operation. In the literature, the term is used in two closely related senses. One usage treats AUCS as the onboard “brain” of an AUV that integrates perception, cognition, decision-making, and adaptive mission planning under uncertainty (Atyabi et al., 2020). Another usage expands AUCS to a mission-driven, multi-vehicle, multi-sensor underwater system spanning ROVs, AUVs, HUVs, USVs, UBVs, underwater communications, docking stations, buoys, and localization infrastructure, with the explicit objective of persistent, adaptive operations across the Smart Ocean (Xu et al., 2024). Later systems instantiate these ideas through cognitively grounded mission frameworks, SLAM-integrated reasoning architectures, distributed ROS 2 agent societies, and dual-brain vision-language-action control stacks (Thierauf, 12 Oct 2025).

1. Conceptual scope and autonomy stratification

The central idea of AUCS is not merely autonomous actuation, but cognitive autonomy: the system continuously ingests heterogeneous sensory data, estimates state and environment, plans energy- and communications-aware actions, executes them with feedback, and updates internal models, maps, or policies (Xu et al., 2024). This broad definition encompasses environmental monitoring, deep-sea exploration, infrastructure inspection, defense, and diver-centric operations.

A recurring distinction in the literature is between autonomy as task execution and autonomy as adaptive reasoning. The survey formulation makes this explicit through levels $L0$–$L5$: $L0$ teleoperation; $L1$ assisted autonomy; $L2$ supervised mission autonomy; $L3$ collaborative autonomy; $L4$ cognitive autonomy with adaptive sensing, model-based decision-making, anomaly detection, autonomous docking and recharge, and network-aware operations; and $L5$ resident, self-sustaining autonomy with underwater docking stations and energy harvesting as a future target (Xu et al., 2024). By contrast, the mission-planning literature argues that coarse “levels of autonomy” are often less informative than explicit allocation of cognitive functions between human and computer across mission phases (Atyabi et al., 2020).

This duality helps clarify a common misconception. AUCS is not restricted to a single monolithic controller on a single vehicle. In one research line it is a layered cognitive subsystem onboard an AUV (Atyabi et al., 2020); in another it is an ecosystem-scale architecture combining vehicles, communication links, support facilities, and edge or shore compute into a persistent ocean service network (Xu et al., 2024). A plausible implication is that AUCS should be interpreted as an architectural class rather than a single implementation template.

2. Architectural principles and cognitive organization

Across the cited work, AUCS is consistently organized as a layered or modular stack separating reactive control, deliberative planning, execution monitoring, and learning. A survey synthesis motivates a layered architecture with a reactive layer for low-level control and collision avoidance, a deliberative layer for task and route planning under multi-objective constraints, an executive layer for monitoring and contingency management, and a learning/adaptation layer for online model learning, policy improvement, and predictive health models (Atyabi et al., 2020).

A more concrete realization appears in DINOS-R, a cognitively grounded, teleo-reactive mission planning and execution framework developed for AUV Sentry (Thierauf, 12 Oct 2025). DINOS-R combines symbolic goals and facts, PDDL-like preconditions and effects specified programmatically in Python, a deliberator that arbitrates among operator goals, safety rules, and mission goals, and behavior implementations that remain separate from their symbolic definitions. Its teleo-reactive rule schema is stated as

$(g_i \land c_i) \Rightarrow a_i$

with execution determined by the highest-priority applicable rule at each cycle (Thierauf, 12 Oct 2025). This architecture is intended to provide understandable, repeatable, and eventually provably safe behavior.

A second architectural lineage integrates cognition directly with navigation and memory. A Soar-based AUCS couples SLAM, attention, working memory, semantic memory, procedural memory, episodic memory, adaptive navigation, and first-order safety reflexes on a shared memory substrate (Jayarathne et al., 14 Nov 2025). Here, the working memory representation mirrors the SLAM state and map, while production rules encode condition–action knowledge such as obstacle-triggered replanning and adaptive sensor management. The design goal is to reduce false loop closures in dynamic scenes and maintain long-term map consistency (Jayarathne et al., 14 Nov 2025).

A third lineage distributes cognition across specialized AI agents. UROSA decomposes AUCS into ROS 2-native agents, including a Commander AI Agent, Perception and Scene Reasoning Agent, Motion Planning Agent, Autonomous Code Synthesis Agent, Predictive Diagnostics Agent, Capability Assessment Agent, and Digital Twin Curator Agent, each wrapped with a Safety Parser (Buchholz et al., 31 Jul 2025). This suggests an AUCS can be centralized in logic yet decentralized in execution, provided interfaces, retrieval-augmented grounding, and safety validation remain explicit.

3. Physical substrate: vehicles, sensing, localization, and communications

The AUCS concept presupposes a heterogeneous vehicle ecosystem. ROVs provide tethered, real-time control and high-precision manipulation, including deep-sea operations with high payloads but with cost, tether logistics, and limited range as constraints. AUVs provide untethered systematic surveys, mapping, environmental monitoring, search and rescue, mine countermeasures, and infrastructure inspection. HUVs bridge tethered and untethered modes or air–sea transitions. USVs serve as relay nodes, edge compute platforms, and deployment/recovery gateways. UBVs provide low-noise, low-drag, energy-efficient motion for sensitive ecosystems or stealth tasks. Gliders provide buoyancy-driven, very energy-efficient long-duration profiling with slow sawtooth trajectories (Xu et al., 2024).

AUCS perception is correspondingly multimodal. Common payloads include multibeam and side-scan sonar for bathymetry and seafloor imagery, DVL and INS for dead reckoning, pressure sensors for depth, CTD and dissolved-oxygen sensors for environmental state, cameras for inspection and biological classification, and magnetometers for anomaly or object detection (Xu et al., 2024). In fielded multi-modal perception systems, these modalities are fused in factor-graph or optimization-based SLAM back ends; for example, a work-class ROV platform with three synchronized cameras, STIM300 IMU, and Teledyne Workhorse Navigator DVL demonstrated real-time reliable state estimation and 3D reconstruction in Trondheim Fjord deployments (Kaveti et al., 6 Jun 2025).

Localization and mapping follow canonical marine formulations. Time-of-flight ranging is given by

$d = c_w t / 2$

for acoustic ranging, while Kalman filter fusion uses the standard prediction and update recursions

$L5$0

and

$L5$1

with SLAM built from sonar, vision, INS, and DVL cues (Xu et al., 2024). The broader mission-planning literature additionally frames belief estimation via EKF, particle filtering, and factor-graph SLAM under nonlinear process and measurement models (Atyabi et al., 2020).

Communications remain one of the defining AUCS bottlenecks. Acoustic communication is the only practical long-range underwater wireless link, with ranges and data rates varying strongly by carrier band: low frequencies support tens to hundreds of kilometers at tens to hundreds of bps, while high frequencies support tens to hundreds of meters and up to $L5$2 kbps. Optical links operate in the blue–green window and can reach up to $L5$3 Mbps at $L5$4 m in clear water but degrade in turbidity. Hybrid acoustic–optical systems use acoustic for long-range control/status and optical for close-range high-rate data, though commercial products are not yet available (Xu et al., 2024). Canonical models include acoustic transmission loss

$L5$5

or

$L5$6

Shannon capacity

$L5$7

and optical attenuation

$L5$8

(Xu et al., 2024).

For AUCS networking, these physical limits force architectural consequences. Underwater sensor networks and UIoT rely on store-and-forward and delay-tolerant operation, with USVs and surface buoys acting as acoustic-to-RF or satellite gateways and performing edge analytics (Xu et al., 2024). In cognitive acoustic networks, POMDP-based scheduling has been shown to maximize end-to-end throughput while constraining primary-user throughput loss; in that setting, time-slot allocation can outperform frequency-slot allocation by up to $L5$9 under certain traffic conditions (Peng et al., 2023).

4. Planning, decision-making, and control under uncertainty

AUCS decision-making operates under partial observability, communication latency, uncertain ocean currents, and strict energy budgets. The mission-planning literature formalizes this through MDP and POMDP models, multi-objective costs of the form

$L0$0

and energy-aware motion models using hydrodynamic drag and relative current velocity (Atyabi et al., 2020). The same literature emphasizes that contingency management, plan repair, rolling-horizon replanning, and resource reallocation are not optional add-ons but core autonomy functions (Atyabi et al., 2020).

DINOS-R offers an explicitly symbolic and teleo-reactive answer to this problem. Missions are specified declaratively as goal formulas plus numeric bindings; the deliberator composes behaviors, inserts calibration or recovery steps when required, and accepts mid-dive operator retasking through acoustic commands (Thierauf, 12 Oct 2025). In AUV Sentry dive 768, a high-level acoustic command changed the goal to “ready for recovery” about 30 minutes before survey end, after which the system re-sequenced recovery behaviors and completed recovery without incident (Thierauf, 12 Oct 2025). This example illustrates a distinctive AUCS property: online plan revision at the mission-logic level rather than only at the trajectory-tracking level.

Other systems push adaptive decision-making into perception-language interfaces. UnderwaterVLA separates a cloud “cortex” for long-horizon chain-of-thought mission decomposition from an onboard “cerebellum” that emits JSON-structured action commands with rationale, discrete direction, and discrete speed level (Wang et al., 26 Sep 2025). These commands are converted into hydrodynamics-informed MPC setpoints. The local execution loop runs at $L0$1 Hz over fixed $L0$2 s motion primitives with quadratic drag compensation and online drag-parameter estimation from IMU-derived accelerations (Wang et al., 26 Sep 2025). In field tests, this dual-brain design reported task completion improvements of $L0$3 to $L0$4 over baseline under degraded visual conditions (Wang et al., 26 Sep 2025).

Control abstractions in AUCS vary by subsystem. The survey reference architecture places PID, MPC, and RL-based controllers behind a common control layer, with the canonical PID law

$L0$5

and energy accounting through

$L0$6

(Xu et al., 2024). In manipulation-specific systems, control can be learned end-to-end. AquaBot maps dual-camera visual features to continuous $L0$7-DoF force/torque commands plus gripper action, then improves beyond human teleoperation through surrogate-based optimization of a time-invariant action-scaling vector $L0$8, achieving a $L0$9 speed improvement over a human operator and $L1$0 over the base behavior-cloned policy in object grasping (Liu et al., 2024).

5. Human–robot interaction, semantic perception, and application domains

Human–robot teaming is an explicit AUCS requirement rather than an afterthought. The broader autonomy literature stresses supervisory control, trusted autonomy, dynamic authority transfer, operator transparency, and explanation as key determinants of adoption and workload (Atyabi et al., 2020). DINOS-R operationalizes this by giving operator goals priority over autonomous mission goals and by allowing insertion or removal of steps both pre-dive and mid-dive (Thierauf, 12 Oct 2025).

Diver interaction is a concrete example of AUCS perception-to-cognition interfacing. An interpretable underwater diver gesture recognition system trained on the CADDY dataset classifies 17 classes, including Start, End, Up, Down, Photo, Mosaic, Boat, Carry, Here, Delimiter, and None, and reached $L1$1 accuracy on a held-out test set of 3093 images with a ResNet18 model (Mangalvedhekar et al., 2023). The cited work explicitly maps these gestures into symbolic commands for state machines or behavior trees, enabling mission toggling, motion adjustment, rendezvous requests, and safety overrides. Integrated Gradients and Occlusion Sensitivity are used to verify that the model attends to diver hands and finger configurations rather than background artifacts (Mangalvedhekar et al., 2023). This is significant for AUCS because it couples semantic intent recognition with auditability.

Semantic perception also supports infrastructure and ecological monitoring. A multi-modal spatial-perception system for a work-class ROV combines multi-camera geometric SLAM, IMU, and DVL with DeepLabv3 semantic segmentation to build semantic 3D point clouds for pipeline, support, and background classes (Kaveti et al., 6 Jun 2025). A different AI-powered underwater exploration system uses YOLOv12 Nano for detection, ResNet50 for feature extraction, PCA retaining $L1$2 variance, K-Means++ clustering, and GPT-4o Mini for structured summaries, reporting $L1$3, precision $L1$4, and recall $L1$5 on a merged 55,722-image marine dataset (Almazrouei et al., 8 Dec 2025). That system does not implement navigation or SLAM, but it exemplifies the AUCS perception–analysis–reporting pathway.

Application domains in the surveyed AUCS literature include environmental monitoring with CTD transects and bathymetry, deep-sea exploration combining large AUV mapping with large ROV manipulation, infrastructure inspection with hybrid ROV–AUV mode switching, mine countermeasures and reconnaissance by AUV swarms, stealth sensing with UBVs, and diver-centric human–robot teaming enabled by wearable localization and gesture interfaces (Xu et al., 2024). Manipulation extends this further: AquaBot demonstrated autonomous object grasping, trash sorting, and rescue retrieval in real-world experiments, including towing a 6.8 kg humanoid object (Liu et al., 2024).

6. Evaluation, safety, limitations, and future directions

AUCS evaluation spans autonomy, mission outcome, resource efficiency, communication quality, localization, endurance, reliability, and human teaming. The survey formulation lists autonomy level achieved, mission success rate and coverage, energy per km and per mapped km$L1$6, communications throughput and latency, localization accuracy, endurance, depth capability, mean time between failures, and dock turnaround time as core metrics (Xu et al., 2024). The mission-planning literature adds situation-awareness quality, contingency-handling success, decision latency, collision or near-miss count, operator workload, trust calibration, and intervention frequency (Atyabi et al., 2020).

Empirical validations reported in the cited papers are heterogeneous. DINOS-R demonstrated simulation and at-sea deployment on AUV Sentry with successful dynamic goal retasking and recovery (Thierauf, 12 Oct 2025). UnderwaterVLA reported improved task completion under degraded visual conditions, but its experiments were tank-based and hardware details were not specified (Wang et al., 26 Sep 2025). UROSA reported $L1$7 diagnostic accuracy on thruster fault cases, $L1$8 collision-free decentralized negotiation with $L1$9–$L2$0 s negotiation times, and autonomous ROS 2 node generation success rates of $L2$1–$L2$2 depending on task, including a dynamically generated DVL+compass filter that reduced drift rate by about $L2$3 compared with DVL dead reckoning after INS failure (Buchholz et al., 31 Jul 2025). AquaBot reported perfect $L2$4 grasping success for its MLP policy in one object-grasping evaluation, with an average completion time of $L2$5 s, and generalization of learned speed scaling to trash sorting and rescue retrieval (Liu et al., 2024).

Several limitations recur across the literature. Energy remains the most persistent systems constraint, especially for long-range AUVs and resident operations (Xu et al., 2024). Acoustic bandwidth and latency constrain real-time cognition, and hybrid acoustic–optical systems, while promising, are not yet commercial (Xu et al., 2024). Long-duration reliability is challenged by corrosion, pressure, component wear, and biofouling (Xu et al., 2024). Environmental domain shift affects both classical perception and learned models through turbidity, lighting variation, and acoustic variability (Xu et al., 2024). Formal verification is still largely aspirational: DINOS-R reports linear temporal logic constraints and PRISM/SPIN integrations as under test rather than complete (Thierauf, 12 Oct 2025), and UROSA explicitly identifies formal verification and real-time performance guarantees as future work (Buchholz et al., 31 Jul 2025).

A final misconception worth addressing is that “cognitive” necessarily implies end-to-end learned autonomy. The literature does not support that narrow view. AUCS implementations range from symbolic deliberators and teleo-reactive rules (Thierauf, 12 Oct 2025), to Soar-based memory and reasoning (Jayarathne et al., 14 Nov 2025), to factor-graph state estimation plus semantics (Kaveti et al., 6 Jun 2025), to behavior-cloned manipulation with online surrogate optimization (Liu et al., 2024), to distributed LLM agents with retrieval and code generation (Buchholz et al., 31 Jul 2025), to dual-brain VLA-plus-MPC systems (Wang et al., 26 Sep 2025). This suggests that AUCS is best understood as a cognitive systems integration problem: combining perception, knowledge representation, adaptive planning, action execution, and safety assurance within the distinctive physical constraints of underwater robotics.