Argo AI

Contribution to AI

The most durable thing to come out of the Ford–Volkswagen bet was a dataset. When Argoverse appeared in June 2019 with a CVPR paper behind it, self-driving corpora offered sensor logs and boxes; this one shipped the map as well — roughly 290 kilometres of lanes with direction, drivable area and ground height, alongside 3D tracking annotations and some 300,000 mined vehicle trajectories from Pittsburgh and Miami. That made map-conditioned motion forecasting a benchmark anyone could enter rather than a capability only fleet operators could evaluate, and the trajectory-prediction literature of the following years was largely scored against it. Argoverse 2, presented at the NeurIPS datasets track in 2021, extended the idea with a thousand annotated sensor sequences, twenty thousand unlabelled lidar sequences, twenty-six object categories and a per-scenario map. Both remain downloadable under a Creative Commons licence, outliving the company that gathered them. The 2017 purchase of Princeton Lightwave carried single-photon Geiger-mode detection at eye-safe wavelengths out of defence imaging and into road vehicles, chasing several hundred metres of range on dark, low-reflectivity objects. A five-year, fifteen-million-dollar research centre at Carnegie Mellon followed in 2019. The dissolution in October 2022, and Ford's $2.7 billion write-down, ended the assumption that deep-pocketed carmakers could simply buy their way to driverless operation.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).