Latency would be kept low be keeping the compute nearby. One 1U or 2U server per robot would be reasonable.