Robust Autonomous UAV Landing on Maritime Platforms via Multimodal Agentic AI and Active Wave Compensation

Technical Summary: Robust Autonomous UAV Landing on Maritime Platforms via Multimodal Agentic AI and Active Wave Compensation

Problem Statement
Autonomous aerial inspection of maritime infrastructure, such as offshore wind turbines and oil rigs, is frequently hindered by the stochastic nature of sea states. These conditions introduce significant risks, including high-kinetic impacts during touchdown, post-landing toppling due to deck inclination, and sensory occlusion caused by rapid platform movement. Traditional strategies often burden the Unmanned Aerial Vehicle (UAV) with the entire task of compensating for wave-induced roll and pitch, a task that becomes unreliable in Global Navigation Satellite System (GNSS)-denied or sensory-degraded environments. Furthermore, reactive stabilization systems like PID or Sliding Mode Control (SMC) often exhibit "chasing behaviors" in dynamic sea states, failing to maintain the level surface required for safe landing.

Methodology
The authors propose a decoupled, multi-vehicle landing framework that synchronizes an Unmanned Surface Vehicle (USV) equipped with a 3-Revolute-Prismatic-Universal (3-RPU) stabilized platform with a robust UAV. The system operates without active cross-platform communication, relying instead on two independent Deep Reinforcement Learning (DRL) agents to manage distinct aspects of the landing process:

1. USV Wave-Motion Compensation (SAC Agent): A model-free DRL agent utilizing the Soft Actor-Critic (SAC) algorithm controls the 3-RPU parallel mechanism on the USV. This agent learns a non-linear mapping between vessel IMU data and platform displacement to actively counteract wave disturbances. The control input consists of a 10-component state vector (including vessel and platform roll/pitch rates and errors) processed through a lightweight neural network to output continuous pitch and roll rate commands. This is complemented by a secondary PID controller for heave stabilization and a centering term to maintain actuator workspace.
2. USV Position Hold: A cascaded PID controller (position → velocity → thrust) uses GPS pose estimation to maintain the USV's position and yaw, ensuring the landing platform remains within the UAV's sensing range.
3. Multimodal Agentic UAV Landing: The UAV employs a behavioral-level agentic policy based on a Transformer-like architecture. Instead of a sequential perception-state-control pipeline, this agent fuses raw visual, thermal, and LiDAR data directly into a latent state representation. A policy network, trained via reinforcement learning, maps this representation directly to high-level navigation actions. This approach ensures resilient navigation under sensor degradation or GNSS denial by avoiding explicit pose estimation.

Key Contributions
The paper outlines three primary contributions:

• Reliable Multi-Vehicle Landing: A mission profile utilizing an independent USV hub for logistics and energy management, maximizing UAV inspection time by ensuring landings occur in a stabilized zone away from hazardous structures.
• Independent Wave Mitigation via DRL: A 3-RPU mechanism controlled by SAC that achieves significant motion reduction (cited as 86.1% in contributions, 87.8% in abstract) to facilitate touchdown without requiring input from the aerial vehicle.
• Resilient Agentic UAV Landing: A multimodal policy that maintains navigation under diverse sensor degradations, achieving a field-proven precision of 0.10 ± 0.08 m without external positioning systems.

Results
The system was evaluated in a high-fidelity Gazebo-based maritime simulation across 15 trials under calm, moderate, and rough sea states (wind speeds up to 6.0 m/s).

• Success Rate: The system achieved a 100% landing success rate across all trials.
• Stabilization Efficacy: The 3-RPU platform demonstrated a mean stabilization efficacy of 87.8% (87.9% in the results section). While the vessel experienced tilts up to 3.06°, the stabilized landing surface (Nest) maintained a mean tilt of 0.33°.
• Horizontal Plane Maintenance: In the roughest conditions, the landing surface remained within 1° of the horizontal plane for 96% of the mission duration.
• Landing Precision: In calm and moderate states, the mean landing error remained under 0.6 m. In rough states, mission time increased (peaking at 97.24 s) as the UAV's policy prioritized safety by awaiting stabilized landing windows.
• Limitations: While the level deck ensured static stability, the GPS-based position hold showed limitations in harsh wave states, leading to increased center-distance errors in specific trials due to drift and multipath interference.

Significance and Claims
The paper claims that by decoupling the mechanical stabilization of the landing deck from the UAV's navigation logic, the system effectively neutralizes the risks of high-kinetic impact and post-landing toppling. The authors assert that mechanical deck stabilization combined with robust, agentic target detection is a critical enabler for safe, objective maritime inspections in hazardous environments. The work suggests that this dual-vehicle approach ensures consistent target perception even when environmental severity would typically induce sensory occlusion or mechanical instability in single-vehicle systems.

Future work identified by the authors includes transitioning from simulation to real-world deployment, replacing GPS-based hold with sensor-fused localization to reduce drift, and implementing bi-directional communication to synchronize control actions.