Lite VLA: Efficient Vision-Language-Action Control on CPU-Bound Edge Robots