“Vision-Language-Action (VLA) models are a "bad hack" for robotics because they are inaccurate, slow, and too expensive for practical applications.”