“A RoboTTT model trained with an 8,000-timestep context outperforms the same model pretrained with a 1,000-timestep context by 62%.”