World-Model-Grounded LLM Planning for AUV and ASV Navigation Near Offshore Wind Farms
Large Language Models (LLMs) can decompose complex mission prompts into robot actions, but lack physical grounding-often ignoring hydrodynamic drag, current drift, or thruster lag. This work pairs an LLM planner with a neural-surrogate world model, a gradient-based trajectory optimizer, and an MPC closed-loop replanner with a trust-region guard. The LLM sets the high-level action sequence, while the world model optimizes action durations to guarantee feasible, collision-free navigation for a 6-DOF AUV and a 3-DOF differential-drive ASV.
In simulated offshore wind farm scenarios, world-model grounding reached a 100% success rate with zero collisions versus 0% for ungrounded LLM plans. In GazeboSim trials with waves and currents, it cut goal-distance error by 70–82% for the ASV and ~93% for the AUV. A companion VLM pipeline also converted satellite maps, nautical charts, and weather forecasts into 96%-accurate obstacle maps for sensor-cheap ASV navigation.
This work has been accepted to the IROS 2026 AQ²UASIM workshop.
SAGE: Ergodic Control for Autonomous and Adaptive Inspection of Subsea Infrastructure
Subsea infrastructure inspection currently relies on fixed schedule routes that treat all components equally, regardless of risk. When a leak or anomaly emerges mid-mission, traditional fixed-tour path planners (A*) cannot adapt without manual operator re-planning. SAGE (Semantic and Adaptive Generative Ergodicity) is an ergodic-control architecture that dynamically allocates AUV time in proportion to a live, sensor-derived risk distribution rather than a scripted route.
By projecting both the risk distribution and vehicle trajectory into Fourier space, the spectral multiscale coverage controller continuously drives down coverage error on every control cycle. In simulated two-XT inspection scenarios, SAGE automatically increased revisit frequency for a high-risk valve-checking it every 5.8s compared to a fixed 8.1s loop under A*-while adapting instantly to new risk weights without explicit re-planning steps or human intervention.
This work has been accepted to the IROS 2026 AQ²UASIM workshop.
A Collaborative Reasoning Framework for Anomaly Diagnostics in Underwater Robotics
AURA (Autonomous Resilience Agent) is an AI framework I developed to help operators manage underwater vehicles when things go wrong. Using Human-in-the-Loop Distillation, it translates noisy sensor data into clear anomaly hypotheses, works with the operator to diagnose root causes, and retains that expertise so the system keeps getting smarter. This creates an adaptive, explainable partner for resilient human-robot teams in safety-critical marine missions. Accepted to IROS 2026, AURA is part of my broader research direction: building AI that does not merely classify sensor readings, but reasons about them and explains its conclusions.
Distributed AI Agents for Cognitive Underwater Robot Autonomy
UROSA (Underwater Robot Self-Organizing Autonomy) is a new software architecture I created to rethink how underwater robots make decisions. Instead of hard-coding every behavior, UROSA uses a team of specialized AI agents that share information, negotiate plans, and adapt collectively as the mission evolves. This is a deliberate move away from brittle, rule-based autonomy toward genuinely cognitive underwater systems that can handle the unexpected without waiting for a human to micromanage them. It combines distributed AI, multi-agent coordination, and robust control into an architecture designed for the unpredictability of the ocean.
Postdoctoral Research Project: Robots UNITE for Offshore Wind Farm Maintenance
This postdoctoral research targets one of the hardest automation problems in the energy sector: maintaining offshore wind farms without sending people to sea. I developed a multi-agent robotic framework that coordinates an Autonomous Underwater Vehicle (AUV), an Autonomous Surface Vehicle (ASV), and a robotic manipulator into a single, cooperative inspection and maintenance system.
The work combines advanced software architectures, autonomous control, and marine dynamics to keep robots operating reliably in harsh offshore conditions. Core technical contributions include dynamic disturbance rejection, tightly coordinated multi-agent control, and real-time decision-making under uncertainty. Beyond the algorithms, the project evaluates the real-world feasibility of uncrewed wind farm maintenance, with the potential to cut operational costs, reduce human risk, and lower the carbon footprint of crewed service vessels.
Motion Control for Tethered Multi-Robot Systems in Marine Environments with Disturbances
Operating a tethered underwater robot from a surface vehicle is deceptively difficult: every wave and current acting on the ASV is transmitted straight to the AUV through the cable. This work introduces a novel motion-control framework that explicitly models and compensates for these coupled disturbances, enabling an Autonomous Underwater Vehicle (AUV) and Autonomous Surface Vehicle (ASV) to operate as a single coordinated system in real marine conditions.
The approach bridges nonlinear marine dynamics, multi-robot coordination, and robust control to keep the tethered team stable and purposeful despite waves, wind, and currents. It is a direct contribution to making tethered marine robotics dependable enough for field operations.
Building an AUV from Scratch - Ocean Systems Lab, Heriot-Watt University
We built a fully autonomous underwater vehicle from the ground up as a research platform - not a kit, not an off-the-shelf ROV, but our own machine: mechanical structure, propulsion, electronics, and a complete sensor suite (DVL, sonar, vision) integrated under a ROS 2 control architecture. Fitted with a robotic manipulator, the vehicle goes beyond inspection into intervention, which is what makes it scientifically interesting: a floating base and a moving arm form a coupled multibody system where every motion of one disturbs the other.
The platform serves as a working testbed for the core problems in underwater autonomy - multibody dynamics and whole-body control of floating-base manipulation, SLAM and perception in turbid, unstructured conditions, and fully autonomous mission execution from planning through to manipulation. Because we control the hardware, software, and control stack, new ideas move quickly from derivation to simulation to open-water validation on a system we understand end to end.
Underwater Manipulation
This research focuses on underwater manipulation, combining context-aware behavior learning with heuristic motion memory. The goal is to push precision and efficiency in real subsea operations by developing motion planners that recover gracefully when sensors fail and compute fast enough to run onboard. Accepted to IROS 2025, this work advances trajectory and full-body motion optimization for underwater vehicle-manipulator systems operating in harsh, partially observable environments.
Teleoperation of an Underwater Vehicle-Manipulator System
A hybrid teleoperation setup for controlling an underwater vehicle and its manipulator together: a joystick drives the AUV, while a webcam tracks the operator's hand in real time to open and close the gripper and set the end-effector position. Hand tracking and depth estimation run through my camera_depth_ros2 pipeline, and the arm motion is planned and executed with ros2_moveit_manipulators using MoveIt 2. The system was tested in a pool with the physical AUV and manipulator before moving to open water.
Underwater Visual Servoing with an Alpha 5 Manipulator on a BlueROV2 Heavy
Building on the teleoperated vehicle-manipulator platform, I achieved closed-loop underwater visual servoing: an Alpha 5 manipulator mounted on a BlueROV2 Heavy autonomously tracks and follows an object detected by the underwater camera, with no operator driving the arm. The vision pipeline continuously estimates the object's position and feeds it to the arm as a moving target; at every control step, the manipulator re-solves its inverse kinematics (IK) against that updated position, so the end-effector's joint trajectory is constantly re-planned to chase the object rather than reach a single fixed point. The result is smooth, reactive motion that keeps the gripper locked onto the target even as it drifts with the current or moves across the camera's field of view.
RAG-Semantic Integration for Adaptive Multiagent Underwater Autonomy
Underwater operations are constrained by limited communication, poor visibility, and unpredictable currents, conditions where traditional teleoperation and rule-based autonomy often fall short. This work bridges high-level reasoning and physical execution by combining Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG), structured Knowledge Graphs, and domain-specific taxonomies. The resulting framework lets heterogeneous AUVs - survey and intervention platforms - dynamically plan, delegate tasks, and share situational data, while a Human-in-the-Loop mode brings the operator back in when the unexpected happens. High-level LLM decisions are grounded into dynamic Behavior Trees, which manage task execution safely in real time and adapt to changing system states such as low battery or sensor noise.
Structured knowledge grounding eliminated LLM hallucination of non-existent capabilities or unsound mission plans, and the framework achieved a 100% mission validation success rate with complete Behavior Tree generation across all tested scenarios. In realistic marine simulations, it demonstrated resilient multi-AUV collaboration, including dynamic task hand-offs, remote data transfer, error correction in object classification, and adaptive survey planning. This work was published in the IEEE Journal of Oceanic Engineering (2026).
URoBench: Comparative Analyses of Underwater Robotics Simulators from a Reinforcement Learning Perspective
This paper introduces URoBench, a modular benchmarking framework to standardize the evaluation of Reinforcement Learning (RL) algorithms across underwater robotics simulators. By rigorously testing HoloOcean, Dave, and Stonefish under identical RL conditions, it demonstrates that fair, comparable, and reproducible evaluation is possible in marine robotics-something the community needs to move learning-based methods from simulation to sea trials.
Modeling and Control of the BlueROV2 Heavy: Dynamics, PID, and Model Predictive Control
Building on the same underwater robotics work, I put together a full technical write-up on modeling and controlling the BlueROV2 Heavy, from first principles to actual hardware commands. It walks through the complete 6-DOF rigid-body and hydrodynamic model of the vehicle, explaining physically what each term means: added mass as displaced water, Coriolis coupling as the same effect that deflects long-range projectiles on a rotating Earth, and restoring forces as the passive righting behavior that keeps the vehicle upright without any controller at all. From there it covers how a desired force and moment gets distributed across the vehicle's eight thrusters, and the often-overlooked final step of converting each thruster force into the PWM pulse the electronics actually expect.
Two control strategies are developed on top of that model: a classical bank of six PID loops, simple and effective for station-keeping, and a Model Predictive Controller that reasons about the coupled, constrained system as a whole and anticipates actuator limits before they're hit. The document is written to be pedagogical rather than just a reference, so a student can rebuild the entire pipeline, from equations of motion to the signal on the wire, rather than treat any part of it as a black box. This is a living document, and it will keep growing as the modeling and control work continues.
This write-up grew out of my postdoctoral research at Heriot-Watt University, and it is meant as a practical resource for understanding the BlueROV2 Heavy's dynamic model together with both PID and MPC control.
Neural Surrogate World Models for LLM-Guided MPC
This one sits outside the underwater robotics work above - it reflects a current interest of mine in World Models rather than a postdoctoral project. Yann LeCun has argued that a genuinely intelligent agent needs an internal world model: a system that can simulate the likely consequences of an action before taking it, rather than reacting to raw observations one step at a time. That idea - plan by imagining, not by acting first and observing the result - is what this project sets out to test in a small, fully checkable setting.
Another one from the same line of work, this time closer to a game than a robot. I built a "think before you act" agent for a maze navigation task, based on the idea that a controller can plan by simulating candidate futures inside a learned model of the world rather than a physics engine. The agent looks a handful of moves ahead, scores each imagined outcome, and only ever commits to the first step of whichever plan looks best, then repeats the whole process at the next timestep. The trick that makes this fast is replacing the actual simulator with a small neural network trained to imitate it: the network's predictions track the real environment closely enough that the difference is barely visible in the numbers, while planning against it runs about seven times faster than planning against the simulator directly.
What made the project satisfying was how little was needed to get there - three possible moves, two numbers read off an image as a reward signal, and a greedy look-ahead search - small enough that every result could be checked by hand, which is exactly what you want before trusting the same idea on something bigger. An LLM sits on top as the high-level planner, querying this world model to score candidate plans before an MPC controller commits to the next action, rather than using the LLM only for next-token prediction. This is very much a foundation rather than a finished result, and the natural next step is seeing how far the same recipe carries into richer, less toy-like environments.
