Current industrial changeover tasks heavily rely on manual programming, leading to significant time consumption and demanding specialized operator skills. This paper proposes a novel framework that leverages multimodal large language models (MLLMs) for initial scene understanding and decision-making, followed by a VLA model for task exploration and trajectory generation to enhance the programming intelligence of cobots. The core mechanism involves an “explore-then-reproduce” strategy: the VLA model autonomously explores and records successful task trajectories during initial setup, which are then rapidly reproduced for subsequent identical changeover batches. Preliminary experiments demonstrate a significant reduction in changeover time by approximately 30% and reduce the programming skill requirements for operators. This work represents the first successful application of large models’ exploration and reproduction mechanisms to enhance the intelligence and versatility of cobots in addressing industrial changeover challenges, offering a new paradigm for flexible manufacturing.