COMCORE: Conflict Resolution for Vessels

Updated 18 February 2026

COLREGs-Compliant Conflict-Resolving (COMCORE) is a framework that uses formal rule encoding, probabilistic risk assessment, and optimization-based planning to resolve multi-vessel conflicts.
It integrates methodologies like state machines, MPC, deep reinforcement learning, and control barrier functions to ensure adherence to COLREGs and prevent collisions.
Extensive simulations and field tests demonstrate that COMCORE achieves high compliance with near-zero collision rates and real-time performance in diverse marine environments.

A COLREGs-Compliant Conflict-Resolving (COMCORE) framework is a class of system architectures, algorithms, and formal methods designed to ensure autonomous marine vessels navigate safely by resolving dynamic multi-vessel conflicts in compliance with the International Regulations for Preventing Collisions at Sea (COLREGs). These systems integrate perception, risk quantification, situational classification, optimized motion planning, and/or learning-based control to guarantee both collision avoidance and regulatory adherence across a range of vessel sizes, encountered geometries, and operational domains. COMCORE methods span classical search, optimization-based MPC, probabilistic reasoning, deep reinforcement learning, control barrier functions, and rely on formal encodings of overtaking, crossing, head-on, and emergency rules as defined in COLREGs.

1. COLREGs Rule Formalization, Situation Assessment, and State Machines

COMCORE approaches universally rely on explicit formalization of COLREGs rules, typically Rules 8, 13–17. These rules are encoded in various mathematical forms: state machines, temporal logic, geometric sectors, or cost terms. Situation assessment is routinely implemented via a multi-state automaton for each vessel–obstacle pair: e.g., SF (safe), OT (overtaking), HO (head-on), GW/SO (give-way/stand-on crossing), EM (emergency) (Bergman et al., 2020, Eriksen et al., 2019). Entry into each state is triggered by thresholds on time to closest point of approach (TCPA), distance at CPA (DCPA), and relative bearings. The assessment may include uncertainty-aware computations where TCPA, DCPA, and encounter geometry are modeled as random variables with Gaussian error structure and corresponding risk/confidence integrals (Hansen et al., 2024).

Rules are then mapped to automata or statecharts, which prioritize emergency responses (highest), then overtaking/head-on/crossing, and finally stand-on maintenance. The transitions among states are governed by observed quantities (bearing, heading, relative velocity), and sometimes, by stochastic evaluation of encounter types and their likelihoods (Hansen et al., 2024, Bergman et al., 2020).

2. Optimized Motion Planning and Model Predictive Control

A central pillar of COMCORE systems is optimization-based motion planning—both single-agent and multi-agent. Typical architectures employ a two-stage pipeline: a global lattice-based path planner producing a sequence of feasible, dynamically-constrained motions and discrete COLREGs states, followed by a local, receding-horizon optimal control problem (OCP) for trajectory refinement (Bergman et al., 2020, Eriksen et al., 2019). The lattice planner constructs a kinodynamically feasible library of motion primitives, enforces infinite penalty for both collision and COLREGs violation, and uses conflict rules to dictate forbidden maneuvers.

The receding-horizon stage fixes the discrete sequence of COLREGs interaction types and solves a direct multiple shooting OCP to smooth the trajectory and optimize for secondary objectives such as energy, path length, or cross-track deviation, all while maintaining explicit separation and regulatory constraints.

Constraints for regulation compliance are embedded as either hard geometric forbidden regions (static/dynamic obstacles, target-ship domains) or as energy-penalizing soft potentials that bias the optimizer’s solution to be proactively compliant, e.g., penalizing motion to the forbidden side in crossing or head-on encounters (Bergman et al., 2020, Vries et al., 2022). The system can be extended by integrating probabilistic margins to robustify against sensing or prediction uncertainty (Bergman et al., 2020, Hansen et al., 2024).

3. Probabilistic, Risk, and Uncertainty-Aware COMCORE Methods

COMCORE systems increasingly leverage explicit modeling of state estimation uncertainty and probabilistic risk. Vessel states (position, heading, speed) for both ownship and contacts are treated as stochastic variables, with fused covariance estimates from sensor fusion or Bayesian filtering (Hansen et al., 2024). Monte Carlo sampling, kernel density estimators, or analytic closed-forms are then used to propagate this uncertainty through TCPA/DCPA and encounter-angle calculations.

Encounter situations (head-on, crossing, overtaking) are evaluated probabilistically, yielding not just a single-label classification but the full posterior over possible rule scenarios for both ownship and target (Hansen et al., 2024). Decision thresholds (e.g., activation of a give-way maneuver) are set according to confidence levels, e.g., triggering only if the probability of “give-way impending collision” exceeds a required α_thr. This enables formal guarantee that rule-compliance and conflict resolution occur with user-defined statistical confidence, and allows “fail-safe” behavior under high uncertainty.

4. Deep Reinforcement Learning and Learning-Based COMCORE Architectures

Many modern COMCORE systems adopt deep reinforcement learning (DRL) to achieve implicit rule-compliance through structured observation and reward design (Meyer et al., 2020, Larsen et al., 2021, Vekinis et al., 2023, Krasowski et al., 2024). State representations combine navigation features (cross-track error, heading, velocities) with context-encoded perception (sectorized range/velocity data, object detections). Action spaces are typically continuous thrust and yaw commands, matching marine actuators.

Reward functions incorporate explicit COLREGs-inspired penalty terms: sector-based distance and velocity penalties distinguish starboard/port/give-way/stand-on logic; trade-off coefficients modulate path-following priority versus obstacle proximity, with the balance tipped automatically in higher-risk scenarios (Meyer et al., 2020, Larsen et al., 2021). Systems with logical action masking enforce compliance with COLREGs—even under RL training—by exposing only those actions which provably satisfy all rule constraints (Krasowski et al., 2024).

Learning-based agents attain near-100% compliance in both simple encounters and realistic, multi-vessel, AIS-derived traffic (Meyer et al., 2020). Stochastic variants have been shown to yield formal lower-bounds on compliance and collision risk reduction (Hansen et al., 2024).

5. Hybrid and Hierarchical COMCORE Frameworks

State-of-the-art COMCORE architectures employ layered designs uniting global path planning, tactical (mid-horizon) model-predictive control (MPC), and short-term, robust collision-avoidance modules, such as branching-course MPC (BC-MPC) (Eriksen et al., 2019, Eriksen et al., 2019). The global layer computes an energy-optimal path among static obstacles. The tactical COLAV layer, executed periodically, perturbs this path to comply with moving obstacles and COLREGs, using explicit state-machine classifiers and regulation-aware constraints or costs in the MPC. The short-term branch acts as a safety net, rapidly sampling emergency maneuvers and enforcing last-resort COLREGs actions, even when other vessels deviate from protocol.

Complementary methods include RRT+Velocity Obstacle hybrids, where a tree of globally feasible waypoints is grown with candidate edges pruned if they violate dynamic regulation-derived forbidden velocity cones (Dubey et al., 2021). Visibility-graph and voxel-based planning is also employed, especially in constrained environments (urban canals, riverine contexts) (Meyers et al., 2022).

6. Distributed, Robust, and Formal Safety Guarantees

Recent extensions have advanced distributed COMCORE architectures using control barrier functions (CBFs), embedding the COLREGs rules as safety barrier sets for each pair of vessels (Paine et al., 16 Jan 2026). At every control cycle, each USV independently solves a quadratic program to minimally modify nominal behavior while guaranteeing forward invariance of both geometric (collision) and regulation-derived (rule) safe sets, even under worst-case, adversarial actions by other agents. Encounters are classified on-board via bearing, and the active rule determines which additional barrier constraints are enforced. This achieves formal, adversarially robust, real-time safety and is communication-light.

Temporal logic encoding (e.g., Metric Temporal Logic) further enables agents to guarantee satisfaction of all regulation constraints under all possible closed-loop trajectories, with the statechart ensuring proper rule priorities (emergency ≻ crossing/head-on/overtake ≻ stand-on) (Krasowski et al., 2024). The action space can be pruned such that only compliant trajectories are ever selected, yielding zero collision and zero violation rates even in adversarial or multi-agent environments.

7. Experimental Validation and Scalability

COMCORE methods have been validated in extensive simulation and full-scale field deployments across diverse platforms and traffic densities (Thakar et al., 2023, Eriksen et al., 2019, Paine et al., 16 Jan 2026). Multi-agent grid-based planners demonstrated real-time performance (sub-100 ms planning) and scalability to fleets of 10+ vessels (Thakar et al., 2023). Real-world tests on urban canals, rivers, and open sea confirm robust conflict resolution, regulatory compliance (no violations or collisions in hundreds of scenarios), and practical run-time (<0.5 s for nonlinear program solves with multiple vessels and obstacles) (Vries et al., 2022, Thakar et al., 2023, Paine et al., 16 Jan 2026). Quantitative metrics commonly include minimum CPA maintained, compliance rates per regulation, time/distance penalties, and CPU time per maneuver.

DRL-based and formal-invariant COMCORE agents have achieved >99% compliance and near-zero collision rates in high-density AIS-derived traffic, even with sensor and prediction noise (Meyer et al., 2020, Hansen et al., 2024, Krasowski et al., 2024).

In summary, the COMCORE paradigm integrates machine-readable regulation encoding, geometric and risk-based situation interpretation, optimization and/or RL-based planning, and rigorous real-time conflict resolution into a unified, scalable, and formally verifiable framework for autonomous surface vessel navigation under the COLREGs (Bergman et al., 2020, Eriksen et al., 2019, Vries et al., 2022, Paine et al., 16 Jan 2026, Hansen et al., 2024, Thakar et al., 2023, Meyer et al., 2020, Agyei et al., 2024).