ASVSim (AirSim for Surface Vehicles): A High-Fidelity Simulation Framework for Autonomous Surface Vehicle Research

1. Problem Statement

The maritime industry is increasingly adopting Unmanned Surface Vehicles (USVs) for inland waterway and port transport to improve efficiency and address labor shortages. However, developing and testing autonomous navigation systems in the real world is hindered by:

• High Costs and Safety Risks: Real-world testing is expensive and dangerous.
• Complexity: Inland waterways feature narrow passages, locks, dense traffic, and sensor occlusions (e.g., from bridges), requiring robust control strategies.
• Lack of Open-Source Tools: Existing simulators are either proprietary (e.g., Kongsberg K-Sim, Nautis), designed for human training rather than algorithm development, or lack the necessary high-fidelity (visual realism and accurate physics) required for training deep learning models and Reinforcement Learning (RL) agents.
• Data Scarcity: There is a critical shortage of open-source, high-quality synthetic datasets for maritime computer vision and sensor fusion.

2. Methodology

The authors introduce ASVSim, an open-source simulation framework built upon Cosys-AirSim and the Unreal Engine 5 (UE5). The framework is designed specifically for Inland Waterway Transport (IWT) and integrates three core dimensions of high fidelity:

A. Vessel Dynamics (Physics)

• Model: Implements a 3-Degree-of-Freedom (3DoF) maneuvering model based on the Fossen formulation. This is standard for inland vessels where vertical dynamics (heave, roll, pitch) are negligible due to calm water conditions.
• Equations: The dynamics account for rigid-body mass, hydrodynamic mass, Coriolis/centripetal forces, and damping (linear and non-linear).
• Disturbances: Explicitly models environmental disturbances, including wind and currents (assumed irrotational and constant over short spans).
• Control: Supports arbitrary thruster configurations (bow, stern, azimuth) with configurable thrust and angles.
• Vessel Models: Includes pre-configured models for the MilliAmpere Ferry, Qiuxin No.5, CyberShip II, and a generic Mariner class cargo vessel.

B. Sensor Simulation

• Radar Emulation: Since standard AirSim LiDAR does not mimic marine radar, the authors developed a Point Spread Function (PSF) based method.
  • It converts dense LiDAR point clouds into radar-like images.
  • It simulates the elliptical resolution cells of pulse-echo marine radars, accounting for range and azimuth beam width.
  • It approximates the behavior of units like the Furuno DRS12A.
• Visual Sensors: Utilizes UE5's photorealistic rendering for cameras, depth sensors, and LiDAR.
• Data Output: Generates ground-truth labels automatically, including semantic segmentation masks, bounding boxes, and depth maps.

C. Environment & API

• Procedural Generation (PCG): Allows for the runtime generation of randomized port and waterway environments (varying channel widths, angles, and obstacles) to support Domain Randomization for RL training.
• APIs: Provides Python and Matlab APIs for vessel control, sensor data retrieval, and environment manipulation.
• Integration: Supports ROS, headless mode (for server-side training), and integration with Cesium for Unreal (for geospatial data).

3. Key Contributions

1. Open-Source High-Fidelity Simulator: ASVSim is the first open-source framework combining photorealistic rendering, Fossen-based vessel dynamics, and marine radar simulation specifically for autonomous shipping.
2. Radar Simulation via PSF: A novel approach to emulate marine radar imagery from LiDAR data, bridging the gap between optical simulation and radar perception research.
3. Synthetic Dataset Generation: The ability to automatically generate large-scale, annotated datasets for object detection, segmentation, and depth estimation, addressing the lack of maritime data.
4. Procedural Environment Generation: A system to create diverse, randomized training scenarios to improve the generalization of RL agents.
5. Comprehensive API: Enables seamless integration with existing control and learning libraries (e.g., PythonRobotics, Stable Baselines3).

4. Experimental Results

The authors validated ASVSim through three primary experiments:

A. Waterway Segmentation (Computer Vision)

• Setup: A pre-trained model (WaSR-T) trained only on real-world data was tested on synthetic ASVSim images without fine-tuning.
• Result: The model achieved high performance on synthetic data:
  • IoU (Water): 0.9916
  • IoU (Air/Sky): 0.9859
  • IoU (Obstacles): 0.9418
• Implication: The simulator's visual fidelity is sufficient for direct transfer of vision models, suggesting potential for pre-training models in simulation before real-world deployment.

B. Path Planning with DWA (Traditional Control)

• Setup: Implemented the Dynamic Window Approach (DWA) for local path planning in a channel with obstacles.
• Result: The vessel successfully navigated a 11-waypoint trajectory without collisions.
• Limitation: The computational cost was high (~1.3 seconds per planning step on a single CPU thread), making it unsuitable for real-time operation without optimization (e.g., GPU acceleration).

C. Reinforcement Learning (RL)

• Setup: Trained an agent using CrossQ (an off-policy actor-critic algorithm) to navigate procedurally generated port channels with static (buoys) and dynamic (other vessels) obstacles.
• Observation Space: 54-dimensional vector (waypoints, heading, velocity, LiDAR data).
• Result:
  • The agent achieved a ~60% success rate during training and maintained a 57.2% success rate on 250 unseen test episodes.
  • Failures were primarily due to obstacles spawning directly in front of the vessel, highlighting the limitations of a single-rear-thruster control scheme.
• Implication: ASVSim effectively supports RL training with domain randomization, enabling agents to generalize across varied environments.

5. Significance and Future Work

• Significance: ASVSim fills a critical gap in the research ecosystem by providing a free, high-fidelity tool that bridges the "Sim-to-Real" gap for inland autonomous shipping. It enables researchers to develop, test, and validate algorithms (both traditional and deep learning-based) safely and cost-effectively.
• Future Directions:
  • System Identification: Calibrating the simulator with real-world data from the authors' physical USV to improve the fidelity of the dynamics model.
  • 6DoF Models: Extending the physics engine to 6 degrees of freedom for open-ocean scenarios.
  • Advanced Sensors: Adding models for multi-beam echosounders and radar beacons (RACONs).
  • Real-World Transfer: Conducting extensive real-world validation to prove the efficacy of algorithms trained purely in ASVSim.

Availability: The framework is open-source under the MIT license and available at https://github.com/BavoLesy/ASVSim.