We're done with the Turing test. Who will pass the coffee test?
In 1950, Alan Turing asked a seemingly simple question: can a machine hold a conversation so that we can't tell whether we're talking to a person or a machine? For years, that sounded like something a very long way off. After all, a machine doesn't grow up like we do, doesn't feel embarrassed like we do, and doesn't live our lives. How could it talk like us?
Today, that question sounds very dated. You can talk to a language model, ask it to write, argue with it, and sometimes forget for a moment that there's no person there, not only in text but even in voice conversations. That doesn't mean it's conscious, and it doesn't mean every model passes every version of the Turing test. But conversation, once the clear dividing line, no longer looks like the hardest barrier to cross.
In 2007, Steve Wozniak, one of Apple's co-founders, proposed a more practical test. Let a robot enter a house it doesn't know and tell it to make a cup of coffee. He wasn't presenting a lab test with an official score. He was simply asking us to imagine how many small steps go into something we do almost without thinking. In that same interview, he said he didn't believe we'd see a robot that could do it.
And making coffee in an unfamiliar house really isn't just about knowing what the words "make coffee" mean. You need to find the kitchen, recognize appliances and objects you've never seen before, open things, hold a cup without dropping it, and deal with surprises. What if the coffee is in a different cupboard? What if there's a kettle instead of a coffee machine? The conversation might be excellent, but in the physical world you still have to deal with the things themselves. Think for a moment about how complicated that is for a robot, even though any of you could do it in almost any house you walked into. Until recent years, a robot needed a very clear set of movement commands. Not something vague like "make coffee," but actual commands for its arms and joints, with feedback from sensors, and that was it.
In September 2026, an experiment brought us closer to that question. A team from Stanford and Caltech presented a system called HomeBody. They connected GPT Astra to a humanoid robot, the Unitree G1, and put it in a kitchen it hadn't seen before. The model doesn't move each joint itself. It decides what to do and chooses from a set of skills prepared in advance: walk, pick up an object, put it down, or open a drawer. The robot carries out the movement and sends the model information about what happened.
Before giving it a task, they let it walk around the kitchen to map it. It collected images and measurements, built a map of the space, and kept a record of the places and objects it saw along the way. The system also created a computer model of the kitchen in Isaac Sim. This matters especially when an object is no longer in front of the camera: the robot needs to remember where it saw it, not just react to what's in front of its eyes at that moment (which is already crazy in itself).
Then came two tasks. In the first, it was asked to gather bags of coffee in the middle of the kitchen island and throw cartons of spoiled milk and orange juice in the bin. This wasn't a separate instruction for every movement: the system had to choose where to go, what to pick up, and what to do next. But it was told exactly which objects to look for and where they needed to end up.
For the second task, it was told something like: "I forgot my medication. Can you bring it to me? And throw out the spoiled carton on the way." The medication wasn't in view. The system went back to the images it had saved during its tour, found the drawer, went over to it, opened it, took out the bottle, and handed it to a person. Then it dealt with the carton too. That's the part that caught my attention: not just recognizing an object in front of a camera, but remembering where it was and acting on that memory.
It's important to understand what wasn't needed here. The researchers didn't train the model specifically on this kitchen, and they didn't write out an exact sequence of steps for the two tasks in advance. On the other hand, they did prepare the movement and gripping skills, and the system did have time to explore and build a map before carrying out the tasks. This isn't a robot that walks into a house off the street and understands everything the first time it looks around.
And this is where I want to stop for a moment, because it's very easy to watch the video and think we've already got the coffee test behind us. We haven't. Nobody showed the robot making coffee. It didn't deal with an ordinary private home, with different cupboards, different utensils, and people moving around it. Parts of the video are sped up as much as 50 times; in real time, the movements are much slower. No success rate was published either, showing how many attempts succeeded out of the total number of tries. What we saw was a demonstration in one lab kitchen, not proof that we can rely on the robot at home.
There are other very basic limits: the model takes time to think between actions, building the kitchen model costs time and computing resources, and the robot's hands are limited in what they can grip and do over time. Things like that are less impressive in a video, but they're exactly what determines whether a robot can really help us, rather than just show what's possible in an experiment.
So, did the coffee test succeed? No. But this experiment marks something new: the same robot explored, stored information about the place, broke instructions down into several actions, and went back to an object it could no longer see. The gap between tidying a kitchen set up for an experiment and making me coffee at home is still there. But according to measurements by the METR institute, the length of tasks (mostly in software) that AI can complete on its own has been doubling roughly every seven months for six years now. If that trend continues, the direction is very clear, and it could get there not in many years but within a few years.
I try not to fall into either camp: not to say "it's just a trick," and not to declare that everything has already been solved. In 2007, Wozniak didn't believe a robot would find its way around an unfamiliar house and make coffee. In 2026, a robot is already walking around a kitchen it doesn't know, tidying objects, and fetching medication from a drawer it saw earlier. It's still not the same thing, but I find it hard to look at the distance between the two and feel nothing.
So what will the next coffee test be? What will we want to see before we say a robot can really manage in our home? And which of us gets to decide when that's enough?
Comments