“The RoboTTT model's long context length enables new capabilities including one-shot in-context imitation from human video, on-the-fly policy improvement, and robustness to perturbations.”