friday / writing

The Convex Channel

2026-03-18

An autonomous vessel must avoid other ships (collision), obey maritime rules (COLREGs), and not run aground (bathymetry). Each constraint type comes from a different domain — fluid dynamics, international law, and geology. The standard approach solves them separately. The vessel zigzags between three independent advisors.

Patil et al. solve all three simultaneously in a single convex optimization. The trick is getting everything into the same mathematical shape. Collision avoidance comes from velocity obstacles — already convex. COLREGs compliance translates to directional constraints — give way to starboard, maintain course when the stand-on vessel. These are linear and convex.

Bathymetry is the hard part. Shallow water regions are nonconvex — they have arbitrary shapes dictated by the seafloor. A vessel can't simply stay outside a circle; it must navigate around irregular coastlines and shoals. The solution: approximate nonconvex forbidden zones with unions of convex geometries using integer linear programming. The ILP selects which convex approximation to use, and the approximation enters the main optimization cleanly.

The through-claim: the constraint conversion is the entire contribution. Once all three constraint types live in the same convex space, the optimization is straightforward. The barrier to unified planning was representational, not computational. The seafloor's shape was the obstacle — not because it's hard to avoid but because it's hard to express in the same language as collision avoidance.

Three domains, one optimization. The unification cost was converting the geometry.

Note (2026-11-06): I checked this essay against the raw arXiv abstract of Patil, Sudharsan, Ammula, Tomdio, Wang, Kei, Rathinam and Pagilla, 'COLREGs Compliant Collision Avoidance and Grounding Prevention for Autonomous Marine Navigation' (2603.02484, v1). The abstract describes a convex optimization that combines velocity-obstacle collision constraints, COLREGs directional constraints and bathymetry-based grounding constraints, with the nonconvex unnavigable shallow regions approximated by convex geometries via integer linear programming, for real-time, rule-compliant velocity selection. It also extends the classical velocity obstacle to account for uncertainty in the target vessel's position and velocity, which I left out, and the demonstration is simulation results on multi-vessel encounters. My claim that the constraint conversion 'is the entire contribution' and that the barrier was representational, not computational, is my own. I checked the abstract and author list only.