I get this question a lot: why bother writing robot programs if there will soon be an AGI brain that does everything from one demonstration?
There are many reasons. For humanoid-like work in very unstructured environments, one general model might be the right shape. For manufacturing it isn't.
Robotics is repeating the LLM story
The hope is the same as with language models: one grand model, a little context, and the robot does the task. A general model does a bit of everything, decently, and that fits a home, where anything can happen. A plant is a known cell and a known part, semi-structured, and it wants a narrow range of tasks done excellently. Most stations don't even need the hardest capability.
Take line feeding. Parts arrive in stillages and have to go onto a conveyor, station after station, almost the same every time. What the plant wants from that station is boring: the same behaviour on the night shift as during acceptance, and an engineer who can verify what the robot will do before it does it.
Boring is the requirement
Simplicity is preferred over complexity, because simple keeps a solution tractable: a program you can audit and explain. When a step fails, you edit the step that failed. With one big model you retrain and hope, and fixing one case can break another.
The EU Machinery Regulation applies from 20 January 2027. Under it, a safety component that uses machine learning for a safety function needs a third-party conformity assessment. Every Vesper program runs inside a certified safety envelope, and is proven in a simulation of the cell before it moves.
Learned is not automatically better
The usual answer is that for complex control, contact-rich steps for example, a VLA will win. In our own applications we haven't found a practical case yet where a VLA beat a combination of representation-based, well-engineered robot programs. Cable insertion is the textbook example, and our first place in the qualifiers of the Intrinsic AI for Industry Challenge came from picking the right tool for each step. Occam's razor applies: when a simpler model does the same job, take the simpler model.
Put the intelligence in the design phase
That is where the AI writes, tests and proves the program. It is also where the real work sits, and people make it sound fancier than it is: hard engineering, productization, and feedback from real deployments, over and over. The robot on the line gets the result of all that and runs it the same way every shift. Execution should be light, boring and predictable.