VLMs as Policies or VLAs?

From Language to Action: How VLMs and LLMs Are Reshaping Robot Control Two paradigms are competing — and converging — in the race to build robots that understand the world. --- For decades, making a robot useful meant writing explicit programs: perceive the scene, plan the motion, execute the action. Then foundation models arrived — systems trained on internet-scale data that can read, reason, and describe the world with remarkable breadth. The question the robotics community has been wrestling with ever since is deceptively simple: how do you take that intelligence and get it into a robot's motors? Two broad answers have emerged, and they reflect a deep philosophical split about what robots actually need. --- The First Answer: Train a Specialist — Vision-Language-Action Models The dominant paradigm of the last few years is the Vision-Language-Action (VLA) model. The recipe is appealing: take a powerful VLM pretrained on billions of image-text pairs, fine-tune it on robot teleoperation data, and let it predict motor commands directly from camera observations and natural language instructions. Models like OpenVLA, RT-2, π0, and π0.5 have demonstrated impressive results across a wide range of manipulation tasks. The promise is end-to-end learning — the model handles perception, reasoning, and action in one forward pass. But this paradigm carries a serious hidden cost. The Catastrophic Forgetting Problem When you fine-tune a VLM on robot data, you risk overwriting the very…