“The RoboTTT robot model and training recipe scales visuomotor context to 8,000 timesteps, which is three orders of magnitude beyond state-of-the-art policies.”