Overview
Sailing performance depends on a continuous interplay between wind angle, sail trim, and boat heading — a control problem that resists simple rule-based solutions. This project builds a physics-grounded reinforcement learning agent capable of finding time-optimal tacking routes between buoys under realistic, stochastic wind fields.
CFD Environment
The aerodynamic backbone is a Lattice-Boltzmann CFD simulation implemented in Wolfram Language. The LBM solver models incompressible 2D flow around a sail cross-section and extracts lift and drag coefficients for a given wind speed, sail angle, and boat-to-wind angle combination.
A full parameter sweep was run across:
- Wind speed and direction variations
- Sail trim angles (0°–90°)
- Boat headings relative to wind (0°–180°)
This produced a 12,000-case dataset of lift/drag coefficients used to train a surrogate aerodynamics model, allowing the RL environment to query forces at any operating point without re-running the CFD solver each step.
Reinforcement Learning Policy
A deep policy network was trained using the CFD-derived aerodynamics model embedded in a sailing environment. The agent must navigate from a start buoy to a finish buoy as fast as possible while managing:
- Tacking decisions — when to change tack to exploit upwind angles
- Sail trim — continuous adjustment of sail angle for peak drive force
- Stochastic wind — wind direction and speed vary according to a random process during each episode
The policy was trained with a time-to-finish reward, learning to exploit optimal VMG (velocity made good) angles and time tacks correctly rather than sailing a straight but slow line.
Results
The trained agent successfully learns to tack upwind at near-optimal VMG angles and handles wind shifts adaptively. Compared to a greedy always-point-at-buoy baseline, the RL policy achieves significantly faster completion times on held-out stochastic wind scenarios.