alpamayo

Why we included this project

Alpamayo 1 takes a different angle on autonomous driving by casting it as a reasoning problem rather than a straight perception-to-action mapping. The roughly 10B vision-language-action transformer takes multi-camera RGB frames, egomotion history, and text commands, then outputs both a natural-language Chain-of-Causation explanation and a 6.4-second future trajectory in the ego vehicle's coordinate frame. That interpretable layer is what makes it worth studying: you get not just predicted waypoints but the reasoning behind them, which helps debug tricky cases like cut-ins or complex intersections. The diffusion-based trajectory decoder and the tight coupling of reasoning text with control output are the parts researchers and robotics engineers are most likely to borrow. One caveat: this repo is now the v1 reference; active development and fine-tuning recipes live in the newer Alpamayo hub, so check there before starting new work.