MODPROD 2025 talk: Automated Scene Generation for Testing COLREGS-Compliance of ASVs

MODPROD is Linköping University's annual workshop on model-based development of cyber-physical systems, where researchers and practitioners from Swedish industry discuss modelling, simulation and systems engineering. In 2025 I gave my first talk there, a few months into the PhD.

What I presented

Random or sampling-based generation and historical traffic data both tend to miss dangerous edge cases. The talk proposed generating diverse, dangerous multi-vessel encounters (3–6 ships) by modelling the COLREGS systematically, and compared two approaches:

• Multi-step refinement (MSR). Functional → logical → concrete scenes. One scene per equivalence class of functional scenarios, solved with NSGA-II.
• Search-based only (SBO). Solve a single numerical problem equivalent to the disjunction of all logical models. Satisfiability is not an issue, but there is no completeness guarantee.

Early results (more than 300 runs, 3–6 vessels, 4-minute timeout)

• Both approaches converged before the timeout in at least 93% of runs.
• SBO was faster for fewer than six vessels. MSR scaled better and was faster at six.
• MSR covered 100% of the equivalence classes at every problem size with an even distribution. SBO's coverage dropped roughly linearly with size.
• Both produced genuinely risky scenes: median TCPA under 9 minutes, and danger-sector risk indices in line with encounters derived from AIS data.

These results fed into the MODELS 2026 paper.