Maritime Autonomous Surface Ship (MASS)

• MASS is a class of waterborne vessels equipped with autonomous perception, navigation, and decision-making systems to operate with minimal human intervention.
• Advanced techniques such as GARSA+DWA, optimal control, and sensor fusion deliver high safety margins and precise COLREGs adherence in dynamic maritime environments.
• Integration of digital twin simulations, real-time collision avoidance, and human–AI collaboration ensures robust system validation and safe certification for autonomous maritime operations.

Maritime Autonomous Surface Ship (MASS) refers to a class of waterborne vessels equipped with advanced perception, navigation, control, and decision-making systems that enable them to operate with reduced or no human intervention. These platforms are designed to execute complex tasks such as route decision-making, collision avoidance, grounding prevention, formation control, and mission adaptation in dynamic maritime environments, including confined waterways, littorals, port basins, and open ocean. The proliferation of MASS research has been driven by the need to enhance navigational safety, improve operational efficiency, and address crew shortages, all while satisfying rigorous international rules such as the COLREGs (Convention on the International Regulations for Preventing Collisions at Sea).

1. Geometric and Optimization-Based Route Planning in Confined Waters

Sophisticated route planning for MASS in restricted or irregularly bounded waters requires explicit integration of geometric, kinematic, and navigational safety criteria. The GARSA (Geometric Analysis-based Route Safety Assessment) framework decomposes waterway boundaries into discrete point features (sharp bends, discrete obstacles) and line features (shore segments), parameterizing any candidate trajectory $P(t)$ via a dynamic width function $w_{P(t)} = \min_{i}\mathrm{dist}(P(t), H_i)$ evaluated against these boundaries. The pathwise envelope of this function, $W(s)$, captures local bottlenecks.

A novel Navigational Safety Index (NSI) quantifies each path's tradeoff between global average width and severity of local narrowings,

$$\text{NSI} = \frac{\bar{W}}{ \sum_i [ -\log_{10}( W(s_i)/\xi ) ] \mathbf{1}_{\{W(s_i)<\xi\}} }$$

where $\xi$ is a threshold for minimal safe waterway width. Candidate paths are scored and $k^* = \arg\max_k \mathrm{NSI}_k$ is selected. Feasibility is enforced through a Dynamic Window Approach (DWA), which, at each control loop, samples velocity/yaw-rate pairs within dynamic constraints and evaluates simulated rollouts with respect to proximity to GARSA waypoints and channel boundaries. This approach demonstrably avoids choke points overlooked by generic graph-based planners, as shown in simulated Hamburg port environments where GARSA+DWA dominates A*-based heuristics in both NSI and empirical collision/grounding avoidance (Xu et al., 11 Jan 2025).

2. Multi-Layered Autonomy Stacks and Real-Time Collision Avoidance

A canonical MASS autonomy stack consists of layered modules for sensor fusion, situational awareness, advisory/control supervision, planning, actuation, and operator interface. In operationalization (e.g., Greenhopper ferry), sensor fusion units (AIS, radar, LiDAR, IMU, marine weather) achieve sub-50 ms latency with triple-modular redundancy and Kalman/particle filter fusion. Situation assessment modules (update rates ≥5 Hz) handle static obstacle mapping, dynamic object tracking, and close-quarters risk scoring.

Collision and grounding avoidance is orchestrated via optimal control-based short-horizon planners (SHP), which solve, at each replan step, a receding-horizon problem minimizing a composite cost:

$$J = \sum_{i=0}^{N-1} [w_1\,C_\mathrm{coll}(x_i) + w_2\,\|p(x_i)-p_\mathrm{ref}(i)\|^2 + w_3\,\|u_i\|^2 ]$$

subject to kinematic and COLREGs constraints. Collision penalties $C_\mathrm{coll}$ are soft exponential functions of predicted minimum approach distance to known and dynamic targets. Rule-aware constraints distinguish between head-on, crossing, and overtaking per COLREGs, dynamically adjusting cost weights and enabling both hard (distance limits) and soft (slack-variable) compliance. Empirically, these architectures have demonstrated 98.5% collision avoidance rates across heavy-traffic Monte Carlo experiments, with strict bounding on planning latency and system availability (Enevoldsen et al., 2023).

3. Regulation-Aware, Uncertainty-Robust Motion Planning

Advances in convex optimization enable real-time trajectory selection that explicitly encodes regulatory compliance, trajectory feasibility, and bathymetric risk. By formulating motion planning as a quadratic program,

$$\min_{v\in\mathbb{R}^2} \|v-v_\mathrm{ref}\|^2, \quad \text{s.t.} \quad Av \leq b$$

where rows of $A$ reflect support vectors for velocity obstacle (VO) constraints (including robustified uncertainty-inflated cones), COLREGs-mandated directional constraints, ILP-approximated bathymetric obstructions (as unions of presampled circles), and strict kinematic speed polytopes, these solvers maintain >550 m separation from obstacles while respecting real-time ($<$6 ms) computational bounds.

Notably, encounter geometry maps directly to constraint side selection, with head-on/crossing invoking starboard restrictions and overtaking suppressing avoidance. This ensures simultaneous adherence to collision avoidance, regulatory requirements, kinodynamic limits, and grounding constraints, as validated in high-fidelity multi-vessel simulation over 10,000 steps (Patil et al., 3 Mar 2026).

4. Learning-Based Perception and Probabilistic Ship Modeling

Robust operation in cluttered or degraded environments necessitates statistically informed perception architectures and nonparametric dynamic models. The MassMIND dataset provides >2,900 LWIR-segmented maritime images covering diverse environmental and scene conditions with class ground truth for sky, water, obstacle, living obstacle, bridge, self, and background. Benchmarked segmentation networks (UNet, PSPNet, DeepLabv3) achieve per-class F1-scores up to 100% (water), with DeepLabv3 performing best overall, particularly in accurately resolving small or low-contrast targets. Early and late fusion with EO imagery, as well as 2.5D occupancy grid representations, bolster the robustness and completeness of the MASS perception stack (Nirgudkar et al., 2022).

For dynamic modeling, nonparametric system identification via ensemble feedforward neural networks enables probabilistic prediction of vessel maneuvering, including epistemic uncertainty flagging for out-of-distribution control scenarios (e.g., tight turning, astern propulsion, strong crosswind). This method iteratively propagates particle rollouts under model samples, quantifies uncertainty via mean and Mahalanobis distance metrics, and demonstrably mitigates performance overestimation in PD-controlled port approaches by exposing system uncertainty envelopes (Wakita et al., 2024).

5. Digital-Twin Frameworks and Simulation Infrastructure

Digital twin (DT) technology is essential for safe, standards-compliant MASS research and deployment. DT capability spans six levels, from offline standalones (Level 0) to fully bidirectional, closed-loop autonomous twins (Level 5). Research implementations (Unity-based, Python/C++/ROS2 middleware) combine high-fidelity 3D hydrodynamic simulation, live streaming of ship state (via AIS, GNSS, LiDAR, IMU, weather), and nested model-predictive safety filters to supervise or audit reinforcement learning policies. These twins enable robust "what-if" scenario replay, automated path optimization under COLREGs and environmental constraints, live trip-resimulation for incident investigation, and accelerated certification (Menges et al., 2024, Menges et al., 2024, Gezer et al., 10 May 2025).

Physical testbeds blend simulation with model-scale or semi/full-scale hardware (e.g., MC-Lab basin, R/V milliAmpere 1, R/V Gunnerus), supporting seamless transitions from digital-only algorithms to full deployment. Froude-similar scaling, synchronized measurement (motion capture, wave probes), and modular vessel firmware ensure consistency, reproducibility, and regulatory traceability across the design and validation pipeline (Gezer et al., 10 May 2025).

6. Human–AI Collaboration, Explainable Decision Support, and Certification

Human–AI collaboration in MASS is underpinned by layered transparency features and explainable decision displays (CPA/TCPA widgets, confidence bands, rule compliance metrics), mapped to operational handover modes (remote supervision, remote control). Human-Unsafe Control Actions (Human-UCAs) concentrate at mode-switch junctures and during critical decision accept/reject loops, particularly under ambiguous or information-overloaded scenarios. Adaptive transparency frameworks, leveraging operator cognitive load and trust metrics, modulate information density and alternative plan presentation to optimize intervention timing and reduce Human-UCA rates (Zhang et al., 19 Sep 2025).

Certification of MASS for international operations follows a staged paradigm—beginning with simulated path-planning validation (≥5,000 nmi, zero near-misses), ascending through hardware-in-the-loop controller bench, crewed/uncrewed voyage trials (up to 50,000 nmi) and culminating in full-scale, IMO-supervised operations. Each phase mandates concrete system availability, fault tolerance, and collision avoidance rates (up to 99.999%), layered redundancy (dual/triple modular controllers), rigorous cybersecurity protocols, and systematic environmental and economic performance auditing. Acceptance requires formal demonstration of COLREGs compliance, reliability, and insurance model alignment (Arnaoot, 2 Apr 2025).

---

MASS thus embodies an overview of advanced perception, safety-centered optimization, probabilistic dynamic modeling, layered autonomy stacks, validated digital-physical simulation, adaptive human interface design, and comprehensive regulatory strategy—collectively enabling the safe, scalable, and economically viable integration of autonomous vessels into the world's maritime operations.