The article highlights a significant disconnect between advances in robotics research and the actual deployment of these technologies in production environments. While recent developments in robotic intelligence and machine learning have resulted in impressive capabilities, such as Vision-Language-Action (VLA) models that can follow natural language instructions to manipulate unfamiliar objects, most robots in industries still operate on fixed routines. For example, industrial robots in automotive manufacturing perform narrowly defined tasks, like welding, but require manual reprogramming for new tasks — a stark contrast to the adaptable systems being showcased in research settings. Despite breakthroughs in areas like simulation-to-real transfer and cross-embodiment generalization, which allow robots to learn and apply skills across different platforms, the deployment of these advanced systems remains limited. In practical applications, such as warehouse bin picking, even learned policies still struggle with unstructured environments. The robots deployed today are often classical systems that lack the flexibility and learning capabilities demonstrated in research. Humanoid robots, though heavily funded and developed, are mostly stuck in pilot phases and require significant human intervention for tasks. The disparity extends to the organizations involved. Robotics research is largely driven by companies focused on breakthroughs in learning algorithms, while deployments are still managed by regional systems integrators using traditional programming methods. This divide means that the innovations seen in labs are not translating into practical, scalable solutions for industries that could benefit from them.