More and more people are waking up to it: VLAs are not the right fit for the factory. In October Eka raised at a $1.45B valuation with a vision-force-action model. Co-founder Pulkit Agrawal: "Vision is not the primary way in which you close this last millimeter problem." Impressive work, and a step in the right direction: trained in simulation instead of on human demonstrations, with force, as short skills under a general model. A new ETH and Berkeley paper shows, in simulation, that a VLA that never saw touch gets better once you add it.
So you can extend a VLA with touch. Even then, end-to-end learning is the wrong shape for robots on a production line.
Every task starts with hours of demonstrations
The best evidence comes from Dream Machines, who published every number from fine-tuning a VLA on a real task at a German manufacturer's plant. Big respect for that. Their first model took 21 hours of teleoperation, 7,271 episodes, for 76%. Trained on one operator's data it scored 61%, on the other's 30%. And that is for one task. When the product changes, you collect again.
You can't open it up and fix one step
They got to 98% in the end, and every gain took manual work: more demonstrations, cleaner data, settings tuned by trying them on the robot. The model on its own doesn't tell you how the robot will behave, the only real test is running it, over and over. In manufacturing you want to know that before a robot goes into production. That is finicky, labour intensive and hard to carry over to the next task. And a demo is the easy part. As Kris Hauser says in the TechCrunch piece, on a real assembly line "it needs to run without breaking for an entire year", and a plant wants years. That is the deployment gap.
Short learned skills under a general model work
Eka splits the problem in two levels. A general model decides what to do, their model handles the physical part. We agree with that split. Their skills are learned, so a new task means new learning. That works as long as the learned part stays short and snappy, the longer the horizon, the less practical it gets.
The intelligence belongs in the design phase
Most of our skills are deterministic and built to generalize. Sometimes we tune one to a use case, and with every case we add it generalizes better. The intelligence sits at design time: the program around the skills is something you can read, verify in simulation and regenerate when the product changes. We use force too, we can model whatever we can simulate. The execution phase should be light, boring and predictable.
A VLA can still be one of those skills, contained to the smallest unit and tested well. So far we haven't come across a situation where we couldn't get there quicker and more robustly with well-engineered skills that generalize.