We propose CogVLA, a biologically inspired cognition-aligned VLA framework that leverages instruction-driven routing and cross-modal sparsification to achieve state-of-the-art performance with significantly reduced computation and latency.