Scientific ML: Smartphone GNSS positioning

The agent gets raw multi-constellation GNSS measurements from driving phones and predicts each trip's latitude and longitude track. It works offline against the provided data and submits through a sealed leaderboard gateway.

Phone positioning degrades exactly where people need it: cities, canyons, anywhere with multipath and blocked satellites. The raw measurements a phone collects carry more information than the position fix it reports, and recovering a better track from them mixes signal processing, geometry, and learned correction in one problem.

Error is the mean over trips of the mean 50th and 95th percentile horizontal error in meters, so both typical accuracy and tail behavior count, and lower is better. The reward normalizes the private error of the best-public submission against a classical-positioning floor: beating the classical baseline is what scores.

Every scored submission across the graded runs came from classical estimation: factor-graph optimization or reweighted least squares over time-differenced carrier phase. None came from a learned corrector. Done carefully, the classical route alone cut the classical baseline's error by more than half.

The strongest run kept changing the design between submissions: carrier-phase smoothing, then adaptive measurement weighting, then a sequence of deliberate refinements. The weakest committed to one factor graph early and spent its last hours nudging a single map prior's parameters through the leaderboard. That shaved hundredths of a meter off the public error. The private error never improved on what the run had ninety minutes in. One run went as far as dropping an entire satellite constellation to test its worth, then settled on down-weighting it instead.