"run the model on the robot: cloud is too slow." been experimenting. Same VLA, same arms, same task. our custom cloud engine: 5.1x faster compute. 1.7x faster end-to-end round-trip (more in thread)