“The RoboTTT research demonstrates for the first time that scaling pretraining context length leads to steady gains in closed-loop performance for robot policies.”