The GPT Moment for Robotics Is Here

Until a few years ago, if you wanted to found a robotics startup, you had to budget for years of work just to get a mechanical arm to move, spend exorbitant sums on custom hardware, and have a team of specialists who looked like they had come straight from NASA. That is no longer the case today. Today, you can have a robot that folds laundry in a real laundry facility, learning from data collected by other robots as well, and the intelligence model that drives it even runs in the cloud, not on a supercomputer hidden in its backpack. Here's the takeaway: We are in the “GPT” era of robotics. Just as language models have democratized AI, robotics is now transitioning from an elite profession to something accessible, scalable, and surprisingly quick to bring to market. You no longer need to be a mechanical engineering genius: today, what matters most is creativity, the ability to collect data, and the desire to integrate hardware, even inexpensive hardware, because the intelligence comes from the model, not the metal. Behind this revolution is a team that looks like it came straight out of a TV show: Quan Vang, co-founder of Physical Intelligence, along with Brian, Chelsea, Sergey, Locky, and Adnan. They left Google X, where they were working on advanced robotics projects, to found a startup with an almost insane mission: to build a model that could control any robot in any environment, for any physically possible task. Quan explains that for years, folding laundry was the “Turing test” of robotics: no traditional algorithm could handle the variety and deformability of fabrics, or the unpredictability of the real world. Then, in two weeks of work with Weave, a startup founded by former Apple employees, they managed to get a robot to actually fold laundry for real customers. Another example: Ultra, a logistics startup, now uses robots to pack Amazon orders in real warehouses, operating with near-total autonomy for hours at a time. Whereas previously each robot was an ivory tower, optimized only for itself, now the models are trained on data from dozens of different platforms. Here's a concrete fact: OpenX, the platform that aggregates data from a fleet of heterogeneous robots, has shown that a “generalist” model outperforms a “specialist” model by 50% on the same tasks. And if you've ever heard that it takes top-tier hardware to run these algorithms, forget it: most Physical Intelligence demos run with the brain in the cloud, and the robot in the field is little more than a smart webcam. This is where the real innovation lies: the problem is no longer “how do I program every single movement?” but “how do I collect the right data and integrate a model that already knows how to operate in different environments?” There is another game-changer that no one tells you about: even identical robots change over time, with minor hardware or software modifications that render the collected data obsolete. Therefore, it is better to train models on a variety of robots, so they learn to handle diversity and become more robust. And this variety already translates today into the ability to perform “zero-shot” tasks, i.e., tasks that do not require specific data collected for that particular task: last year, it took hundreds of hours of training, but now the robot is able to generalize. However, there is one aspect that few people consider: the real barrier for a robotics startup is no longer technology, but integration into the actual workflow. Quan reiterates this point: it is important to understand where robots truly make a difference – often in tedious, repetitive tasks, where a few mistakes are acceptable, and where you can start with human supervision and then automate more and more. The upfront cost has dropped: cheaper hardware, open-source models (Physical Intelligence has released PI0 and PI05 with the same weights as the model used internally), and the ability to test and improve in real-world environments. We are at the beginning of a true “Cambrian explosion” of vertical startups: just as the personal computer led to a proliferation of tech companies in the 1980s, now every sector – from logistics to cleaning, from food service to care – can have its own robotics startup, built by small, agile teams that no longer have to reinvent everything from scratch. But beware: the challenge is not just technological; it also involves the product and the business. It is essential to understand the customer's real needs, adapt to existing workflows, and aim for a rapid break-even point—i.e., achieve economic viability with just a few robots before scaling up. And the contrarian view? Everyone expects the revolution to come from home robots, but the real boom could start with invisible industrial tasks – such as order picking or managing micro-logistics – where there is a tolerance for error and demand is huge. Get ready: tomorrow, the key question won't be “Which robot can I buy?” but “Which workflow can I improve by integrating an intelligent model that learns from everyone?” The key takeaway: Robotics is no longer about sophisticated hardware, but about data, models, and creative integration. If, after reading this story, you've realized that your way of thinking about robotics startups has changed, you can indicate that on Lara Notes by using I'm In – it's not a 'like'; it's the gesture of someone who has embraced a new perspective and identifies with it. And if tomorrow you tell someone why folding laundry was the real “Turing test” of robotics, or how Ultra is revolutionizing logistics, you can use Shared Offline on Lara Notes to tag those who were present: that way, that conversation really counts. This Note comes from Y Combinator and has saved you over an hour and a half of listening time.