When I hear that robotics is approaching its “ChatGPT moment,” I find myself thinking about the customer who has to put a robot to work. At GCCEI, a South Korean startup support organization, I sourced startups for possible trials with corporate partners. That work is the background I bring to one question: what would persuade a company to move from a demonstration to a purchase?

In robotics, I don’t think “the breakthrough” names one clear event. A compelling demo would make the technology easier to imagine. A robot that earns its keep on one production task would establish something else, and a system that transfers reliably to unfamiliar work would go further still. I’m trying to understand which of those milestones the current excitement points toward.

One approach I’m following is vision-language-action models, or VLAs. They use images and language to produce actions that control a robot. The appeal is that a new task might require less programming of individual movements. Google DeepMind’s explanation of Gemini Robotics

To think through what that means for a customer, I’ll use an imagined factory job: placing a part in a tray. I can see the value in a robot that understands the instruction, but the buyer also needs it to locate and grip the part, notice if it shifts and recover when placement fails. How much of that work can a general model take on, and how much still belongs to the installation?

The comparison with Waymo

I use autonomous driving as a comparison because it has a visible history of moving from research to a paid service. A closer look at its milestones, though, makes me cautious about turning that history into a forecast for factories.

Google’s self-driving project began in 2009. Waymo introduced a commercial service in Phoenix in December 2018, opened a fully driverless service there in October 2020 and removed its San Francisco waitlist in June 2024. Each date changed where or how people could use the service. Which one should count as its arrival? 2018 launch; 2020 driverless opening; 2024 San Francisco opening

The factory comparison appeals to me because operators can control more of the environment. For the part-and-tray task, they can arrange for parts to arrive in a known position and keep people out of the robot’s workstation. That may make a useful installation possible before the model can handle much broader variation.

But I can’t estimate a factory deployment date by subtracting a few years from Waymo’s history. Handling identical rigid parts is different from handling flexible packaging. The customer’s task determines which failures matter, and a system can be useful for one job while still struggling with another.

Integration with the production line

If I imagine evaluating the part-and-tray robot for a pilot, the grasp is only where my questions begin. How does the robot know a part is ready? Which tray belongs to the current order? What tells the next machine to proceed?

Factories already coordinate that work through machine controllers, production-management software and business systems, and a VLA has to fit into those operations. If it misplaces the part, something also has to detect the failure and decide whether to retry, stop or ask a person for help.

So I’m interested in what an installation team can reuse. Holding the workstation steady and engineering around known failures may be enough to make one job economical. A model that can handle more variation could reduce that engineering, but I need evidence from the next task or site to know how much carries over.

Training data raises the same question. A photograph can show a completed placement without recording the movements or forces that produced it. Robot demonstrations connect observations to actions and outcomes, but collecting them is physical work. Before I treat a successful pilot as evidence of a repeatable product, I want to know how much new data the next machine needs.

Robotics partnerships and acquisitions

The same installation questions shape how I read the Samsung and SoftBank announcements. What interests me is what an industrial robotics business could bring to an AI company: existing customer relationships and the ability to service machines.

Samsung announced it would raise its stake in Rainbow Robotics to 35%, joining its AI and software ambitions with an existing robotics company. With its October 2025 agreement to buy ABB’s robotics business for $5.375 billion, SoftBank sought a substantial industrial operation alongside its AI investments. At this article’s February 2026 cutoff, the ABB transaction was an agreement, not a completed acquisition. Samsung announcement; SoftBank agreement

I can see why those combinations might help a model reach customers. But the announcements don’t yet tell me how reliably a robot will work on a new production line, or whether the next installation will need less custom engineering. Those are the operating results I’m watching for.

What I would want to learn from a pilot

Back to the purchase question: I would want a pilot to show how the robot handles ordinary interruptions, not only whether it can place the part once. These are the questions I would bring:

  • How long does it keep completing the task before someone intervenes?
  • What happens after a failed grasp or an unexpected obstruction?
  • How much new data and engineering does the next workstation require?
  • What does a completed task cost once supervision, downtime and maintenance are included?

To me, a customer adding a second workstation because the first one works reliably enough would be a more useful sign of adoption than another proposed date for robotics’ “ChatGPT moment.” It would still be evidence about a particular job. But it would answer the question that keeps me following the field: when does a promising capability become something a company wants to keep using?