Ray hit the wall most mobility researchers know by heart, and then built a way around it. The developer behind the Ray0202 GitHub account Ray0202/MR-Traj released a two-stage system that manufactures plausible movement traces instead of waiting for the real ones cities refuse to share.
The impasse is older than any single tool. Privacy law and plain caution keep most GPS location histories under lock, so the models meant to ease traffic and plan transit train on scraps. Ray explained, “This coarse-to-fine design is intended to preserve both global mobility structure and local movement patterns,” pointing to how the split keeps the broad shape of a trip and the local jitter that makes it believable Ray0202/MR-Traj.
Ray argued the coarse-to-fine split is the whole point. The paper behind the code, posted to arXiv arXiv cs.LG, says single-shot generators flatten the small, local detail that makes a trajectory convincing. Its authors claim MR-Traj matches existing methods on overall distribution while pulling ahead on fine-resolution mobility, and that injecting randomness at more than one scale makes it harder to re-link a synthetic path to a real commuter.
But hard limits remain. The framework still leans on whatever real data it learns from, and synthetic movement can drift from reality in ways that quietly bias a downstream model. A city planned on warped traces could end up chasing problems that don’t exist.
The open question is whether planners will trust synthetic paths enough to plan a city on them. Ray’s release suggests the answer is creeping forward, one careful layer at a time — the same open-code momentum zBrandco has tracked as shared research repos reshape how AI work spreads LoRA Powers 98.4% of Single-PEFT Fine-Tuning Models, Hugging Face Data Shows.
