AquaChat: Conversational AI for Aquatic Systems

• AquaChat is a conversational AI framework that converts natural language into symbolic plans, enabling aquatic robotic control and sensor integration.
• It employs a multi-layered architecture that translates free-form commands into parameterized control sequences and real-time vehicle actuation.
• The system is validated across diverse applications, from aquaculture net-pen inspections to water-quality advisory and interactive oceanographic installations.

Searching arXiv for the AquaChat-related papers and neighboring work to ground the article. AquaChat is most directly the name of an LLM-guided ROV framework for adaptive inspection of aquaculture net pens, proposed as a multi-layered system that translates natural-language commands into symbolic plans, parameterized control sequences, and low-level vehicle actuation with real-time feedback and event-triggered replanning (Akram et al., 19 Jul 2025). Across adjacent literature, the same label is also reused for related systems in AUV piloting, water-quality advisory, transparent oceanographic query answering, forecasting interfaces, and interactive water-based installations. This suggests a broader usage in which “AquaChat” denotes conversational AI systems that couple language interfaces to aquatic sensing, simulation, control, or environmental interpretation (Yang et al., 2023, Srivastava, 2018, Lourens et al., 2018, Chen et al., 2 Nov 2025).

1. Canonical formulation in aquaculture net-pen inspection

The canonical AquaChat formulation is the aquaculture inspection framework introduced as "AquaChat: An LLM-Guided ROV Framework for Adaptive Inspection of Aquaculture Net Pens" (Akram et al., 19 Jul 2025). It is built as a three-layered architecture. The high-level planning layer uses GPT-4 and a templated prompt that concatenates the user instruction $q$, environment description, ROV specifications, and inspection constraints to generate a symbolic inspection plan $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$. The symbolic substrate is a custom PDDL schema, aquachat_inspection, with predicates such as system_ready(), region_detected(?area), environment_stable, navigated(?rov), inspected(?area), captured(?area), feedback_received, and replan_needed, and actions including plan, move_to, inspect, capture, replan, and report. The planning step is written as

$$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$

where $K$ is the contextual knowledge set (Akram et al., 19 Jul 2025).

The mid-level task manager maps each symbolic action to parameterized ROV API calls, while the low-level motion layer executes navigation and inspection. Within this formulation, move_to is defined so that the controlled state converges to a desired pose, inspect follows a helical trajectory,

$$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$

and capture records image $I_t$ at state $s_t$ (Akram et al., 19 Jul 2025). Validation was performed both in simulation and in a controlled aquatic setting: the simulated environment used ROS/Gazebo with UUV Simulator, BlueROV2 Simulator, and a custom aquaculture net pen, while the pool experiment used a Blueye Pro ROV X and a net pen measuring $2.13\,\mathrm{m}\times 3.35\,\mathrm{m}$ (Akram et al., 19 Jul 2025).

This canonical formulation establishes AquaChat as a robotics system rather than merely a chatbot. Language serves as an entry point for task synthesis, and the resulting plan is embedded in a closed-loop control stack rather than returned only as text.

2. Recurring architectural pattern across AquaChat-like systems

A recurring architectural decomposition appears across the literature. In OceanChat, recast in the supplied material as an AquaChat-style AUV piloting system, the runtime is organized into an LLM Translator, a Task Planner, and a Motion Planner, joined by an event-triggered replanning loop and driven by real-time Lagrangian data streams (Yang et al., 2023). AquaChat++ extends the aquaculture inspection setting to multiple ROVs with a high-level Natural-Language-to-Symbolic planner, a mid-level Task Manager that allocates and schedules inspection jobs under battery and thruster-fault constraints, a low-level control layer with fault-tolerant thruster allocation, and a real-time feedback loop for event-based replanning (Saad et al., 6 Aug 2025). In the Water Advisor design later summarized as AquaChat, the system is decomposed into Sensor Data Ingestion, Knowledge Representation & Storage, Reasoning & Goal Management, Dialogue Management, and a Presentation & Multi-modal UI Controller (Srivastava, 2018). OceanAI follows the same pattern in a data-services context, combining an instruction-tuned LLM with function-calling, an API Gateway & Dispatcher, a Data Processing Pipeline, a Visualization Module, and a Frontend Chat UI (Chen et al., 2 Nov 2025).

These systems differ in domain and execution substrate, but they converge on a common engineering logic. Natural language is first transformed into an intermediate representation: symbolic goals, PDDL actions, API calls, or function invocations. That intermediate representation is then grounded against a robot, a sensor network, or an authoritative data source. This suggests that AquaChat is less a single implementation than a recurrent systems pattern for bridging free-form language with aquatic-world execution.

3. Planning, control, replanning, and empirical performance

In the single-ROV AquaChat framework, the low-level controller uses the kinematic model

$$\dot x=v\cos\theta,\quad \dot y=v\sin\theta,\quad \dot z=v_z,\quad \dot\theta=\omega,$$

with PID laws of the form

$$u_i(t)=k_p e_i(t)+k_i\int_0^t e_i(\tau)\,d\tau+k_d\dot e_i(t),$$

applied to errors in $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$0, $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$1, $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$2, and $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$3 (Akram et al., 19 Jul 2025). OceanChat/AquaChat generalizes motion planning to current-aware execution by augmenting dynamics with a Lagrangian ocean-current field $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$4,

$\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$5

and it triggers replanning when

$\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$6

or when logical predicates are violated (Yang et al., 2023). AquaChat++ adds a 4-DOF cascaded PID controller and a real-time convex QP for thruster allocation,

$\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$7

where $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$8 models thruster degradation (Saad et al., 6 Aug 2025).

System	Setting	Reported results
AquaChat	ROS/Gazebo plus pool experiment	Plan generation time $\mathcal{P}=\{a_1,a_2,\ldots,a_n\}$9–$$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$0 s; LLM planner $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$1 on structured/unstructured commands; “Move To” RMS position error $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$2 m and yaw error $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$3 rad; “Inspect” helix within $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$4 m
OceanChat/AquaChat	HoloEco canyon-navigation task	LLM-Only: $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$5, $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$6 s; LLM + Motion: $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$7, $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$8 s; AquaChat: $$\mathcal{P}^*=\arg\max_{\mathcal{P}} P_{\mathrm{LLM}}(\mathcal{P}\mid q,K),$$9, $K$0 s
AquaChat++	ROS-Gazebo UUV Simulator with two BlueROV2 vehicles	PSR $K$1, EXESR $K$2; $K$3 improved by $K$4; position errors $K$5 cm and yaw error under $K$6 rad

The empirical record shows a consistent trade-off. LLM-mediated systems improve task flexibility, constraint handling, and robustness under replanning, but they also introduce nontrivial planning latency. In the primary AquaChat paper, this appears explicitly in the comparison between rule-based planning at approximately $K$7 ms and LLM-based planning at $K$8–$K$9 s (Akram et al., 19 Jul 2025). AquaChat++ addresses a different axis of realism by incorporating battery thresholds, thruster-fault feasibility, and multi-ROV coordination, thereby shifting AquaChat from single-vehicle symbolic planning toward resource-aware fleet management (Saad et al., 6 Aug 2025).

4. Water-quality advisory and transparent oceanographic query systems

AquaChat also appears in the literature as a water-information interface rather than a robotic controller. In the Water Advisor specification summarized as AquaChat, water quality is represented as a feature vector

$$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$0

regulatory exceedance is encoded by $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$1, and usage-specific risk is computed as

$$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$2

The conversational layer is formalized as a POMDP over environment state, user goal, and dialogue history, with actions such as ask-clarify(location), fetch-data, advise-safe, advise-unsafe, and show-map (Srivastava, 2018). The design couples sensor ingestion, ontology-backed regulation lookup, and multi-modal presentation, making AquaChat a goal-directed agent for water-quality monitoring rather than only a QA front end.

A South African implementation operationalized a lighter-weight variant through a RASA-based Natural Language Understanding and Dialogue Manager pipeline. It classified intents such as water_quality, water_availability, greeting, and farewell, extracted entities including location and situational variables, and queried or locally mirrored data from the Department of Water Affairs, WESSA, and Cyanolakes (Lourens et al., 2018). In the Cape Town use-case, average precision, recall, and F1 were $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$3, $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$4, and $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$5 before interactive augmentation, and $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$6, $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$7, and $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$8 after adding approximately $$x(t)=r\cos(\omega t),\quad y(t)=r\sin(\omega t),\quad z(t)=z_0-v_z t,$$9 dialogues. Class-level F1 values included $I_t$0 for utter_water_availability, $I_t$1 for utter_water_quality, and $I_t$2 for utter_beach_quality (Lourens et al., 2018).

OceanAI extends the same conversational paradigm into verifiable oceanographic analytics. Each query triggers real-time API calls against NOAA streams, and the pipeline identifies, parses, and synthesizes relevant datasets into natural-language responses and visualizations (Chen et al., 2 Nov 2025). In a blind comparison with three widely used AI chat-interface products, only OceanAI produced NOAA-sourced values with original data references (Chen et al., 2 Nov 2025). This branch of the AquaChat lineage emphasizes provenance, parameterized data access, and reproducibility, replacing symbolic vehicle actions with function-calling over authoritative observational archives.

5. Forecasting, hydrologic modeling, and simulation-oriented extensions

AquaChat-like interaction has also been extended to modeling and forecasting workflows. AQUAH is described as an end-to-end language-based agent for hydrologic modeling, beginning with a natural-language prompt and proceeding through a Context Parser Agent, Dataset Retriever Agent, Perceptor Agent, OutletSelector Agent, ParamInitializer Agent, Operator Agent, ReportWriter Agent, and Feedback Reflector Agent (Yan et al., 4 Aug 2025). Its execution layer runs the CREST hydrologic model with governing equations including

$I_t$3

and

$I_t$4

Cold-start experiments over U.S. basins were blind-scored by hydrologists and GPT-o3 on Model Completeness, Simulation Results, Reasonableness, and Clarity; among the listed models, claude-4-opus achieved the highest average score, $I_t$5 (Yan et al., 4 Aug 2025).

AquaCast contributes a forecasting backend that the supplied material explicitly wraps into an AquaChat interface. The model is a multi-input, multi-output Transformer encoder with a single-layer MLP decoder, jointly embedding endogenous history, rain history, and optional rain forecasts, while forecasting only endogenous variables (Abdollahinejad et al., 11 Sep 2025). On the LausanneCity dataset with forecast horizon $I_t$6, AquaCast (RainFull) achieved approximately $I_t$7 MSE versus $I_t$8 MSE for NoRain, described as a $I_t$9 reduction; on SynthLow, median MSE improved from approximately $s_t$0 to approximately $s_t$1 under RainFull (Abdollahinejad et al., 11 Sep 2025). The paper also gives a concrete conversational wrapper in which a query is parsed, time-series histories and rain forecasts are retrieved, standardized, passed through the model, and re-expressed in natural language (Abdollahinejad et al., 11 Sep 2025).

ChatSim moves the same paradigm into underwater simulation. Natural-language commands are mapped by ChatGPT into a constrained library of function calls executed inside Blender-based OysterSim, and the renderer produces RGB images and segmentation masks for downstream training (Palnitkar et al., 2023). Over $s_t$2 randomly sampled prompts, ChatSim achieved $s_t$3 exact-match to the intended function-call sequence. A U-Net trained on $s_t$4 ChatSim-generated images obtained $s_t$5 mean IoU on a small real-world test set, compared with $s_t$6 IoU for a baseline trained on uniform synthetic data (Palnitkar et al., 2023). In this setting, AquaChat-like language grounding is used not to operate a physical system directly, but to configure the experimental environment in which aquatic autonomy is developed.

6. Ecological narrative, affect, and interactive media

Several adjacent systems replace instrument control with ecological narrative and affective mediation. OceanChat, a web-based conversational platform with animated marine creatures, compared Static Scientific Information, Static Character Narrative, and Conversational Character Narrative in a between-subjects experiment with $s_t$7 (Pataranutaporn et al., 5 Feb 2025). The Conversational Character Narrative condition significantly increased behavioral intentions and sustainable choice preferences relative to static approaches, and the beluga whale character showed stronger emotional engagement across measures including perceived anthropomorphism and empathy. At the same time, deeper measures such as climate policy support and psychological distance showed limited movement (Pataranutaporn et al., 5 Feb 2025). This positions conversational aquatic agents as interventions for pro-environmental behavior rather than only as information systems.

Sensorium Arc goes further by personifying the ocean as a poetic speaker in a real-time multimodal interactive AI agent system built on a modular multi-agent system and retrieval-augmented LLM framework (Bissell et al., 20 Nov 2025). The system enables natural spoken conversations from the ocean’s perspective, blends scientific insight with ecological poetics, and uses keyword detection and semantic parsing to trigger data visualizations and audiovisual playback based on time, location, and thematic cues (Bissell et al., 20 Nov 2025). Developed with the Center for the Study of the Force Majeure and inspired by Newton Harrison’s eco-aesthetic philosophy, it proposes a conversational mode of access to high-dimensional environmental data in which ocean data are treated as a living narrative rather than an abstract dataset (Bissell et al., 20 Nov 2025).

Whispering Water, explicitly recast in the supplied description as an AquaChat prototype, materializes human-AI dialogue through cymatic patterns on water (Wang et al., 26 Jan 2026). The installation stages a four-phase ritual of confession, contemplation, response, and release; routes speech to ASR and emotion2vec+; invokes six large-LLMs in parallel; synthesizes their replies through ElevenLabs; and converts those replies through STFT decomposition, Bark-scale remapping, inverse FFT, and six subwoofer channels into interfering wavefields on the water surface (Wang et al., 26 Jan 2026). No formal user-study metrics are reported, and the work is presented as an exploratory art installation (Wang et al., 26 Jan 2026).

Taken together, the literature shows AquaChat as a convergent label for systems that bind aquatic domains to conversational AI through grounded execution. In robotics, the grounding target is vehicle planning and control; in advisory systems, it is sensor and regulatory data; in modeling, it is simulation and forecast infrastructure; and in eco-aesthetic work, it is narrative, embodiment, and affect. This suggests that AquaChat is best understood not as a single platform, but as an evolving design space for language-mediated interaction with water, oceans, and aquatic infrastructures.