Dojo

Dojo trains a robot skill from one sentence. Type “put the red block in the green bowl” and it builds a MuJoCo scene for an SO-101 arm, collects expert demos, fine-tunes a vision-language-action policy, tests it on scenes it never saw and saves it to a skill library that later prompts build on. An optimizer loop keeps changing the training recipe and keeps only changes that raise the held-out score: on one task it took success on 100 fresh scenes from 67% to 74% (improved, not yet significant). Meta’s Model API plans the task and critiques failed runs, and Subquadratic’s SubQ reads every rollout and past round at once to pick what to reuse. Everything runs in simulation on one Mac.