GeoCoTDrive grounds planning-critical 2D regions first, then retrieves localized 3D priors, improving safety-critical planning metrics on nuScenes, NAVSIM and Bench2Drive
The work proposes GeoCoTDrive, an explicit geometric chain-of-thought framework for vision-language-action driving: it first autoregressively grounds 2D regions tied to ego decisions, then samples localized 3D geometry tokens from a frozen geometric foundation model inside those regions and interleaves them into the autoregressive context for trajectory generation; it also formulates planning-relevant grounding and builds the 146K-annotation PlanningGrounding dataset, improving safety-critical planning metrics on nuScenes, NAVSIM and Bench2Drive.