The engineering cost of tactile imagination: Three hard constraints for VLA deployment seen through UniTac

I noticed an interesting detail: The title of UniTac's paper by UBTECH uses the term "tactile imagination," but when flight control engineers see the word "imagination," their first reaction is concern about real-time performance. After years of working on IMU fusion at DJI, I know all too well that if sensor data arrives just 1 millisecond late, the flight controller has to spend an extra 10 milliseconds compensating, which ultimately results in gimbal image jitter. So when I read that UniTac enables VLA (Vision-Language-Action) models to predict contact states via tactile signals, my biggest concern wasn't how smart the model is, but whether this "imagination" can actually run on real robots.