A curated, research-oriented reading list for unmanned aerial vehicle (UAV) landing on unmanned surface vehicle (USV) in waves, wind, and open-water conditions — with a focus on perception + estimation + control + learning.
Goal: help researchers quickly map the space, compare assumptions, and find practical baselines, code, and open problems.
- Scope
- How to Use This List
- Core Papers
- Simulation, Benchmarks, and Evaluation
- Open Problems
- Contributing
Included
- UAV landing / touchdown on moving USVs (waves, vessel tilt/roll/pitch, wind)
- Timing policies / “land at the right time & place”
- Relative pose estimation, onboard sensing, limited/no-comm settings
- Control: MPC, robust MPC, distributed MPC, safety filters, learning residuals / GP / RL (when relevant)
Excluded (unless directly tied to USV landing)
- Generic UAV landing on static pads
- UAV landing on UGVs without marine motion modeling
Each paper entry is written to be comparable, not just a link.
Legend
- Setting: open water? waves? wind? indoor emulation?
- Sensing: AprilTag / vision / VIO / RTK / IMU fusion / etc.
- Method: MPC / distributed MPC / GP / RL / safety filter / hybrid
- Assumptions: comms? shared states? known deck motion model?
- Validation: sim / hardware / sea trials
- Notes: 1–2 lines on what’s actually new + key limitations
Distributed Model Predictive Control for Cooperative Multirotor Landing on Uncrewed Surface Vessel in Waves
- Link: https://arxiv.org/abs/2402.10399
- Authors / Venue / Year: Jess Stephenson, Nathan T. Duncan, Melissa Greeff — ICUAS 2024 (also on arXiv), pp. 645–651, DOI: 10.1109/ICUAS60882.2024.10557042
- Setting: cooperative UAV–USV landing in waves; choose safer location & timing
- Sensing / State: assumes state measurements available for both UAV and USV (focus is coordination/control, not perception)
- Method: Sequential distributed MPC with artificial intermediate goals + coupling cost; uses a spatio-temporal wave model to reduce severe tilt at touchdown
- Assumptions: bilateral communication; each agent shares only the artificial goal (not full state/control), but measurements exist (and comms may be delayed)
- Validation: simulation study (no hardware in this paper)
- Key idea (1-liner): Coordinate landing time + place using distributed MPC while communicating only a lightweight “goal” variable.
- Notes:
- Strength: clean “minimal-sharing” coordination mechanism that’s easy to compare against centralized MPC
- Limitation: depends on reliable state availability and a modeled wave/tilt representation
- Tags:
distributed-MPC,cooperative-landing,waves,timing
A Time and Place to Land: Online Learning-Based Distributed MPC for Multirotor Landing on Surface Vessel in Waves
- Link: https://arxiv.org/abs/2410.21674
- Authors / Venue / Year: Jess Stephenson, William S. Stewart, Melissa Greeff — ICUAS 2025, pp. 193–199, DOI: 10.1109/ICUAS65942.2025.11007817 (also on arXiv)
- Setting: wave-like conditions; emphasizes timing/place to reduce tilt risk
- Sensing / State: indoor experiments with external motion capture (Vicon) for state; tilt motion emulated by a custom platform attached to a UGV
- Method: distributed MPC + tilt-related coupling cost learned online as a Gaussian Process (GP)
- Assumptions: agents optimize locally and share only artificial goals; communication limitations motivate reduced sharing
- Validation: indoor experimental validation with a tilting rig (UGV-mounted platform), plus comparisons to non-tilt-aware baselines
- Key idea (1-liner): Learn a tilt penalty online (GP) so distributed MPC becomes “tilt-smart” in real time.
- Notes:
- Strength: bridges “distributed MPC theory” to a real experimental setup with measurable success-rate gains
- Limitation: still not full open-water sea trials; perception is not the main contribution
- Tags:
distributed-MPC,GP,learning-cost,tilt-aware,experiments
- Link: https://arxiv.org/abs/2301.00255
- Authors / Venue / Year: Parakh M. Gupta, Èric Pairet, Tiago P. Nascimento, Martin Saska — IEEE Robotics and Automation Letters (RA-L), Vol. 8(2), pp. 744–751, 2023, DOI: 10.1109/LRA.2022.3231831
- Setting: harsh open-water-like conditions (waves + strong wind); goal is touchdown during near-zero tilt windows
- Sensing / State: onboard monocular vision with AprilTag on landing platform; no communication with the USV; estimator predicts vessel motion online
- Method: MPC with (i) a novel objective and (ii) online decomposition of oscillatory vessel motion (FFT + prediction) to time the landing
- Assumptions: no UAV–USV comms; relies on fiducial detection availability and a prediction horizon that is short enough for real-time control
- Validation: simulations + “realistic” tests described in the paper (vision-rate variability, estimator behavior, touchdown tilt statistics)
- Key idea (1-liner): Predict deck motion from onboard vision and commit to landing when the deck tilt is predicted to be small.
- Notes:
- Strength: very clear “timing policy” story implemented inside an MPC loop with realistic sensing constraints
- Limitation: vision dropouts / motion blur remain a major practical risk for open-water operations
- Tags:
MPC,vision,AprilTag,no-comms,open-waters,wind
This section includes direct ship/USV landing RL and a few high-quality moving-platform RL baselines that are commonly reused for deck-landing research.
- Link: https://arxiv.org/abs/2209.08381
- Authors / Venue / Year: Vishnu Saj, Bochan Lee, Dileep Kalathil, Moble Benedict — arXiv preprint, 2022 (DOI: 10.48550/arXiv.2209.08381)
- Setting: VTOL ship landing with 6-DoF deck motion + wind gusts (sub-scale test platform)
- Sensing / State: monocular camera only; tracks a horizon reference bar on the landing platform (pilot-inspired visual cue)
- Method: robust deep RL controller (compared against a nonlinear PID baseline) + a dedicated vision pipeline for relative localization
- Assumptions: relies on consistent visibility of the horizon reference bar; learning targets robustness to gusts
- Validation: Gazebo simulation + real-world tests using a Parrot ANAFI quadrotor and a sub-scale ship platform
- Key idea (1-liner): A “pilot-like” visual feature + robust RL yields stronger ship-landing performance than PID under gusts.
- Notes:
- Strength: rare combo of vision-only + real hardware in an RL ship-landing setup
- Limitation: custom visual reference (horizon bar) may not transfer directly to arbitrary USV decks
- Tags:
RL,vision-only,ship-landing,gust-robustness,real-world
Deep Reinforcement Learning for sim-to-real policy transfer of VTOL-UAVs offshore docking operations
- Link: https://www.sciencedirect.com/science/article/pii/S1568494624006173
- Authors / Venue / Year: Ali M. Ali, Aryaman Gupta, Hashim A. Hashim — Applied Soft Computing, Vol. 162, article 111843, 2024, DOI: 10.1016/j.asoc.2024.111843
- Setting: VTOL-UAV landing/docking on an offshore platform with randomized wave-like disturbances
- Sensing / State: simulation state (paper focus is RL training + sim2real-style robustness, not a new perception stack)
- Method: split landing into approach phase (model-based control) + landing phase (DRL); tests DQN and PPO; uses JONSWAP wave spectra for episode randomization
- Assumptions: sim disturbance model captures enough “wave reality” for transfer; approach controller is available
- Validation: numerical simulation experiments (sim-to-real is addressed via disturbance randomization methodology)
- Key idea (1-liner): Use wave-spectrum randomization + phased architecture to make docking policies more transferable.
- Notes:
- Strength: very explicit sim2real mindset (randomization via maritime wave spectrum)
- Limitation: doesn’t provide a full perception pipeline for onboard-only landing
- Tags:
RL,sim2real,offshore-docking,PPO,JONSWAP
- Link: https://doi.org/10.1007/s10514-024-10162-8
- Authors / Venue / Year: Pascal Goldschmid, Aamir Ahmad — Autonomous Robots (Springer), Vol. 48(4), article 13, 2024, DOI: **10.1007/s10514-024-10162-8
- Code: https://github.com/robot-perception-group/rl_multi_rotor_landing
- Setting: moving platform landing (not specifically marine), with realistic disturbances and a curriculum-style training pipeline
- Sensing / State: ROS/Gazebo training setup; real-world deployment supported (see repo + released datasets)
- Method: RL training with a sequential curriculum; end-to-end system integration in ROS (sim + deployment tooling)
- Assumptions: platform motion is observable in the training setup; primary focus is reproducible training + evaluation
- Validation: simulation + real flight experiments (logs/rosbags/videos provided by authors)
- Key idea (1-liner): A practical, reproducible RL landing stack that many “deck landing” projects can reuse as a baseline.
- Notes:
- Strength: arguably one of the most engineer-friendly RL landing repos (ROS/Gazebo + deployment workflow)
- Limitation: not marine by default — you still need to add wave/tilt and maritime sensing failures
- Tags:
RL,moving-platform,baseline,ROS,Gazebo,reproducible
- Link: https://doi.org/10.1007/s10846-018-0891-8
- Authors / Venue / Year: Alejandro Rodriguez-Ramos, Carlos Sampedro, Hriday Bavle, Paloma de la Puente, Pascual Campoy — Journal of Intelligent & Robotic Systems, Vol. 93(1–2), pp. 351–366, 2019, DOI: 10.1007/s10846-018-0891-8
- Setting: moving platform landing (general case; often cited as a canonical deep-RL landing reference)
- Sensing / State: platform-relative state (paper focus is deep RL control for landing)
- Method: deep RL approach for autonomous landing on a moving platform
- Assumptions: platform motion stays within trained regimes; perception is not the core novelty
- Validation: experiments as reported in the paper
- Key idea (1-liner): One of the earlier widely-cited deep-RL formulations of “land on a moving target”.
- Notes:
- Strength: good “classic reference” for framing RL landing as a control problem
- Limitation: not tailored to wave-induced tilt / open-water disturbances
- Tags:
RL,moving-platform,classic-reference
- Link: https://www.imavs.org/papers/2025/7.pdf
- Authors / Venue / Year: Cansu Yıkılmaz, Christophe De Wagter — IMAV 2025 (16th International Micro Air Vehicle Conference and Competition), paper IMAV2025-7
- Setting: landing a hybrid UAV (Variable Skew Quad Plane) on a moving ship
- Sensing / State: uses simulated dynamics + real ship motion data for validation (paper emphasizes guidance robustness)
- Method: PPO outputs optimal acceleration inputs that feed an inner INDI-based controller; reward shaped around touchdown velocity, deviation, and duration
- Assumptions: inner controller handles attitude/low-level stabilization; RL focuses on guidance/acceleration commands
- Validation: simulations with randomized sinusoidal ship motion + validation using real ship data (no sea-trial flight experiments in this paper)
- Key idea (1-liner): Treat RL as an outer-loop guidance policy that “waits for the moment” while an inner controller keeps the vehicle stable.
- Notes:
- Strength: clean separation of concerns (RL guidance + strong inner-loop control)
- Limitation: real ship data ≠ real flight on a ship; perception stack not addressed
- Tags:
RL,PPO,guidance,ship-landing,hybrid-UAV
- Success rate (touchdown without bounce / slip / crash)
- Touchdown relative velocity (vertical + lateral)
- Tilt at touchdown (roll/pitch, or deck normal angle)
- Tracking error (relative position/heading)
- Time-to-land / energy
- Robustness sweeps: wind levels, wave states, comms delay/dropout, perception failure rate
- VRX (Virtual RobotX) Simulation — Gazebo-based USV competition-grade marine environment (waves, tasks, sensors).
- Gazebo Sim — the core simulator stack used by many marine + aerial robotics projects.
- UUV Simulator — Gazebo/ROS plugins for underwater/ocean environments (useful for ocean-current / fluid plugins even if you only need “water physics”).
- PX4 SITL + Gazebo (gz) — common research baseline for multirotors/VTOL; supports wind and rich worlds.
- PX4 Gazebo Models & Worlds — curated vehicles/worlds used by PX4’s Gazebo Sim workflows.
- PX4-SITL Gazebo Classic Plugins — classic plugin suite (still cited in older papers).
- RotorS Simulator — classic academic multirotor dynamics + Gazebo plugins (includes wind plugin; widely referenced in literature).
- JSBSim (official site) — flight dynamics model (FDM) used in flight simulation research.
- AirSim (docs) — Unreal/Unity-based sim with PX4/ArduPilot SITL support; popular for vision-based UAV research.
- Flightmare (project page) — modular quadrotor sim (Unity renderer + physics engine); common in learning/perception pipelines.
- NVIDIA Isaac Sim (docs) — high-fidelity robotics sim + synthetic data generation.
- ArduPilot SITL (official docs) — widely used SITL workflow (often paired with Mission Planner / MAVProxy).
- PX4 Simulation Overview — entry point for all PX4 sim backends.
- ROS 2 Docs — standard middleware for research pipelines.
- MAVROS — MAVLink ↔ ROS bridge (ROS1/ROS2 branches).
- MAVSDK (guide) — high-level MAVLink API (C++/Python/etc.) used for offboard control + integration tests.
- AprilTag 3 — core AprilTag detector library.
- apriltag_ros — ROS wrapper for AprilTag 3.
- OpenCV ArUco module — alternative marker family + pose estimation tools (often used when AprilTag isn’t required).
- Perception failures: marker loss, glare, motion blur, sea spray
- Latency & comms: delayed state sharing in distributed MPC; packet loss
- Contact dynamics: touchdown friction, hooks/magnets, deck compliance
- Generalization: from indoor tilt rigs → real sea states
- Safety guarantees: hard constraints with learning in the loop
- Benchmarking: lack of standardized wave/wind suites + public datasets
PRs are welcome! Please add entries using the template below.
#### Paper Title
- **Link**: (arXiv / DOI / publisher)
- **Authors / Venue / Year**:
- **Setting**:
- **Sensing / State**:
- **Method**:
- **Assumptions**:
- **Validation**:
- **Key idea (1-liner)**:
- **Notes**:
- Strength:
- Limitation:
- **Tags**: