Scaling VLA robotics requires optimized inference, heterogeneous compute and real-time control on edge hardware.