“The RoboTTT model scales visuomotor context without increasing inference latency.”