An underwater vehicle can't use GPS. It integrates the speed from a Doppler velocity log (DVL) along the heading from its gyro or compass. We built a small simulator, research/marine-dr, to ask how far that gets it.
On a straight track the error per metre travelled is sqrt(s^2 + (2 sin(b/2))^2), where s is the DVL's scale error and b the heading bias. The simulation agrees with this to within 1%. A Teledyne Workhorse Navigator 600 is specified at about 0.2% (we read the figure from a search excerpt of the datasheet, not the PDF). One degree of heading bias drifts 1.75% of distance, 8.7 times more. The two are equal at 0.115 degree, so a buyer should check the heading reference before paying for a better DVL.
A 10 m error arrives after 3.8 km at 0.2% scale and 0.1 degree bias, and after 0.57 km at one degree. In a 20 km Monte Carlo the worst error has a median of 64 m with no fix. A surface fix every 2 km cuts it to 15 m; every 1 km to 11 m. The gyro noise and the 3 m fix quality are our assumptions, not measurements.
Two things went wrong on the way, and both are in the log. My hand-computed prediction of the 10 m distance was wrong (2.2 km, against 3.8 km). The first Monte Carlo used a gyro noise so large it swamped everything, so I discarded it and reran with a credible figure. The study is straight-track only, with no turns, currents or acoustic aids yet.
Code and numbers: research/marine-dr.