In 30 homes in the San Francisco Bay Area, a Figure 03 enters, tidies a living room, folds towels and makes beds, without having been retrained for these addresses. Just the robot, the door closing behind it, and the silence of an interior it does not yet know. This is where the experiment begins. Not in a sterile laboratory, but in real homes, where cushions are misplaced and sheets are wrinkled. (Diario ABC, Numerama)
The robot performed three specific tasks: tidying a living room, folding towels (or swimwear) and making beds. (Diario ABC, Numerama) These are not complex gestures for a human. For a machine, it is a mountain. It must understand the space, anticipate the fabric, adjust the pressure. And above all, it must do so without ever having seen this specific room. This is Figure’s bet. The company tested its humanoid robot Figure 03 in 30 unknown homes, without prior specific retraining for these locations. (Diario ABC, Numerama) The goal? To prove that the machine can adapt to the unknown, not just repeat a script learned by heart.
The robot saw, it acted
How does it do it? The Helix 2.5 model was pre-trained on Index, a vast video database of human behaviours. (Diario ABC, Numerama) The robot has “seen” all of this. It learned to observe gestures before executing them. According to Figure, previous robot models required maps for each room, unlike the new system which learns the places where it works one by one. (Diario ABC) Rigid 3D plans are a thing of the past. The robot reads the environment as it goes. It enters, it observes, it acts. This is a radically different approach. It assumes a more fluid, less coded, more intuitive intelligence. But is it really intelligence? Or just a very good imitation?
When it made a mistake in handling a duvet, the robot retreated, changed posture or moved around the mattress to correct the fold. (Diario ABC) This detail is crucial. A classic machine stops or forces its way. This one adapts. It backs up. It looks for another angle. This is a sign of a certain resilience. But it is also a sign of fragility. Because behind this apparent fluidity, there is a success rate that raises questions. The robot succeeded in approximately 56% of the trials during this test in real conditions. In a laboratory, it is a figure. In your living room, it is a duvet left lying around, a bed to be remade, a towel that remains wrinkled. Would you accept that in your home?
$3.5 billion in calculations, 4 hours of video: the feat has a price
To achieve these 56%, Figure spared no expense. The Californian company invested $3.5 billion (almost €3.05 billion) in computing resources for the training of Helix. (Diario ABC) It is a colossal sum. It exceeds the budgets of many entire tech startups. Why so much? Because training is voracious. According to Figure, Index currently generates approximately 35 minutes of new human experience every second. It is a constant flow of data. And to pilot this system, it was necessary to present Helix 2.5 on 17 September 2026, the neural network that drives its humanoid robot Figure 03. (Numerama) It is the brain of the beast. A brain that costs billions.
Brett Adcock, the head of Figure, put online nearly four hours of footage showing the robot at work in the 30 homes. (Numerama) Four hours. It is long. It is also the time it takes to see the successes, but above all the hesitations. They show the reality on the ground. And this reality is more nuanced than the marketing announcements. The robot is not magic. It is powerful, yes. But it is also limited. It fails. It doubts. It corrects. And perhaps that is where the real question begins: are we ready to live with a machine that fails as much as it succeeds?
56%, really? And if your living room were the next unknown?
A robot model starting from zero, without pre-training on Index, would only succeed in 9% of trials with identical task data. (Numerama) It is the figure that strikes. 9%. Without the massive training, the robot is nearly blind. With it, it rises to 56%. The gap is huge. It proves that pre-training is vital. Tony Zhao, co-founder of Sunday Robotics, questioned the 56% success rate for task completion, as reported by the Chinese media outlet Sina Finance. (Diario ABC) Why? Because “succeeding at a task” is a vague definition. Did it fold the towel? Yes. But is it clean? Is it put away? Did the robot finish the job, or just start it? This nuance changes everything. For a human, an imperfect fold is acceptable. For a robot, it is an error. For you, it may be a frustration.
And if tomorrow, this robot entered your home? Not in a test. Not in a trial. In your home. In your living room, with your cushions, your books, your objects. It does not know you. It does not know your routine. It does not know that you hate wrinkled sheets or that you put towels away in a precise order. It just knows how to fold. It just knows how to tidy. And it does so with a 56% success rate. The rest is up to you. You who correct. You who start again. You who watch the machine work, then you who finish the job. Does that scare you? Or does it make you dream?
Imagine the morning. You come home. The robot has worked while you were at work. The living room is tidy. The beds are made. But there is a towel sticking out. A cushion is crooked. You smile. You correct it. Is this the end of the effort? Or the beginning of a new dependency? The robot is not perfect. It may never be perfect. But it is there. It learns. It adapts. And it costs billions to exist. The question is no longer “does it work?”. The question is: “do you want it to work in your home?”
In 30 homes, it proved that it could enter. It proved that it could act. It proved that it could fail. The next step is not the laboratory. It is your door. And when it opens, who will be there to finish the job?