Accepted · OCEANS 2026 IEEE

Large Language Model Guided Reinforcement Learning for Maritime Ship Collision Avoidance

COLREGs reasoning meets learned vessel control

Yasmin Pang, Arsyi Aziz, Alex Zongo, Peng Wei, Kyle Crandall

Mechanical and Aerospace Engineering · Computer Science · The George Washington University

Preprint Cite

From navigation rules to maneuver decisions.

Language-based maneuver intent guides a learned controller, with supervisory arbitration connecting the two.

COLREGs-guided collision avoidance in a crossing encounter (schematic)A target vessel crosses from the ownship's starboard side. The language model reads the encounter and proposes giving way with a starboard turn; the PPO controller executes heading and speed changes, and an arbiter keeps fallback authority. The ownship passes astern of the target and resumes its route. LLM intent crossing, target to starboard: give way, turn to starboard LLM: COLREGs reasoningmaneuver intent PPO policyheading and speed arbiter ownship command
Schematic. In a crossing encounter with a target to starboard, the language model proposes giving way with a starboard turn; the PPO controller executes it and the arbiter keeps fallback authority.
less route-completion time
Up to 54%
less cumulative risk exposure
Up to 37%
collision-free completed episodes
300 / 300

Time and risk compare the learned and rule-based planners. Episode completion comes from a separate full-pipeline evaluation; neither establishes a general safety guarantee.

01 / Architecture

Separate intent from execution.

Interpret

An LLM supplies COLREGs-based maneuver intent.

Control

PPO proposes heading and speed actions.

Arbitrate

A supervisor combines guidance and control, retaining fallback authority.

02 / Planner comparison

Efficiency gains, with tighter separation.

The learned planner improves time and risk metrics, but maintains smaller minimum separation than the baseline.

Four plots compare path length, travel time, minimum separation, and risk exposure across two-, three-, and four-ship encounters.
Planner performance by ship count. Figure 4
Inspect all 22 encounter trajectories

Ownship trajectories are compared against scripted target vessels.

Trajectories for 22 Imazu cases, showing target ships, the CORALL baseline, and the RL-controlled ownship.
Qualitative trajectory comparison. Figure 3

03 / Full-pipeline evaluation

Fallback control matters.

LLM reliability and episode outcomes (Table I)
CaseUsable intentStrict accuracyEpisode success
164.9%100.0%100.0%
679.5%91.9%100.0%
1873.6%96.6%100.0%
Overall72.7%96.0%100.0%

9,507 LLM calls; strict accuracy uses the high-confidence, single-label subset. Success uses 100 episodes per case.

04 / Limits

Simulation evidence, not deployment certification.

One controlled ownship encounters scripted traffic. Broader geometries, more reliable guidance, and multi-agent control remain future work.

Reference

Cite this work

Pang, Y., Aziz, A., Zongo, A., Wei, P., and Crandall, K. (2026). "Large Language Model Guided Reinforcement Learning for Maritime Ship Collision Avoidance" OCEANS 2026 IEEE, Monterey, CA, USA.

BibTeX 
@misc{pang2026maritime,
  title = {Large Language Model Guided Reinforcement Learning for Maritime Ship Collision Avoidance},
  author = {Pang, Yasmin and Aziz, Arsyi and Zongo, Alex and Wei, Peng and Crandall, Kyle},
  year = {2026},
  note = {Preprint; accepted for OCEANS 2026 IEEE},
  url = {https://bpb-us-w2.wpmucdn.com/web.seas.gwu.edu/dist/9/15/files/2026/09/Yasmin_LLM_Guided_RL_COLAV_OCEANS_2026.pdf}
}