Autopilot

lastmod 2026-06-23created 2026-03-2710 edits

Autopilot (自動運航) refers to a device or system that controls an aircraft so that its course is kept constant without human operation. The same term is usually used for ships as well, but when a distinction is needed, terms such as 'automatic navigation' are sometimes used.

More broadly, the technology for general means of transportation is called autonomous driving, and when the term is interpreted narrowly, it sometimes means only self-driving cars, the technology for automobiles.

Core concepts

Autonomous navigation is broadly a combination of three capabilities.

Perception — sensing the surrounding environment with sensors (radar, LiDAR, cameras, AIS, etc.)
Decision-making — deciding on actions suited to the situation (collision avoidance, route optimization)
Control — operating the actual equipment (engine, rudder, etc.)

Levels of autonomy (IMO criteria)

The International Maritime Organization (IMO) divides ship autonomy into 4 levels.

Level	Name	Description
Level 1	Automation support	Decisions made by humans, some tasks automated
Level 2	Remote control	Crew on board, can be operated remotely
Level 3	Unmanned remote control	No crew, remote operation
Level 4	Fully autonomous operation	AI makes all decisions

Key technology components

Navigation AI — route planning, collision avoidance algorithms (COLREGs compliance)
Sensor fusion — integrating radar + LiDAR + camera + GPS data
Korean: 디지털 트윈 — simulation in a virtual environment
V2X communication — real-time communication with other ships and port infrastructure
Cybersecurity — defending unmanned systems against hacking

Machine learning / deep learning SOTA

A summary of the main tasks where deep learning is applied in autonomous ship navigation and the latest research trends (2024–2025).

Ship detection (Object Detection)

Hybrid architectures combining the YOLO family and Transformer are mainstream. Research on models specialized for optical images and for SAR (Synthetic Aperture Radar) images respectively is active.

Main models

YOLO11 — The latest version of the YOLO series. Overall improvements to the backbone, neck, and detection head. Regarded as one of the SOTA models for ship detection
YOLO-SEA (2025) — Based on YOLOv8. Combines a SESA (SimAM + SENetV2) attention module + an improved BiFPN structure. Achieves [email protected] +5.8% and [email protected]:0.95 +7.2% over the baseline model
SwinYOLOv7 — Combines YOLOv7 + Swin Transformer. 96.59% accuracy on SAR images with 640×640 input. Outperforms 18 existing models
YOLOv11/v12 + Attention (2025) — Integrates BRA (Bi-Level Routing Attention) + Swin Transformer + CBAM. Achieves [email protected] of 98.0% on the SAR Ship Dataset
RT-DETR — Transformer-based end-to-end detection. Accuracy and real-time performance validated in various scenarios including optical, SAR, and infrared

Main datasets

SSDD (SAR Ship Detection Dataset)
HRSID (High-Resolution SAR Images for Ship Detection)
SeaShips
LEVIR-Ship
MSAR

Ship trajectory prediction (Trajectory Prediction)

The main architecture is a Transformer + LSTM/GRU combination that processes AIS (Automatic Identification System) data as time series.

Main research

Transformer-based generative prediction — Learns the distribution of historical trajectories to generate future trajectories. Strong at capturing long-range temporal dependencies
Multi-factor LSTM (2025) — Simultaneously models multiple factors from AIS such as latitude, longitude, COG (course over ground), and SOG (speed over ground)
CNN + DNN + LSTM + GRU ensemble — Study based on Korean AIS data (PLOS One, 2024). Characterized by preprocessing that integrates contextual information such as outlier removal, environmental conditions, and ship characteristics

Collision avoidance (Collision Avoidance)

Deep reinforcement learning (Deep Reinforcement Learning, DRL) is the core methodology, and a key challenge is including COLREGs (International Regulations for Preventing Collisions at Sea) compliance as a constraint.

Method	Features
DRL (PPO / SAC)	Continuous action space, handles multi-ship scenarios
MADDPG	Multi-agent cooperative avoidance
Federated DL	Privacy-preserving distributed training
MPC + DL hybrid	Combines physical constraint compliance + neural network decisions

A 2024 study improved both decision-making stability and the COLREGs compliance rate with an entropy-optimized RL algorithm. Companies such as ABB and Kongsberg are building rule verification modules that monitor AI-planned routes in real time into their autonomous navigation systems.

Sensor fusion (Multi-Sensor Fusion)

Real-time fusion of radar, LiDAR, AIS, and cameras greatly improves detection reliability compared to a single sensor. According to a 2025 study, combining radar + AIS + video inputs improved target detection reliability by more than 90%, and AI-based fusion methods can reduce tracking error by about 99% even when AIS data is unavailable.

Route optimization and weather routing

In 2024 field studies, AI-based weather routing consistently achieved 5–10% shorter voyage times and milder sea conditions compared to standard routes. A case was also reported in which, using a predictive model, decisions about worsening weather were made 36 hours earlier than human navigators.

SOTA summary by task

Task	Main models	Core techniques
Ship detection (optical)	YOLO11, RT-DETR	Attention, BiFPN
Ship detection (SAR)	SwinYOLOv7, AC-YOLO	Swin Transformer
Trajectory prediction	Transformer + LSTM	AIS time-series modeling
Collision avoidance	DRL (PPO / SAC)	COLREGs-constrained reinforcement learning
Sensor fusion	Multi-modal Fusion Net	Radar + LiDAR + AIS + Camera
Route optimization	GNN + RL	Combined weather and fuel optimization

Autonomous cargo ships

Favorite site

Autopilot - Namuwiki

author Zer0 sitemap