From the outside, there is nothing special about the building. No bright sign. No big company logo. Just a small door in San Francisco with a tiny symbol on it that looks like the math sign for pi. That simple sign is the only clue that you have reached the headquarters of Physical Intelligence, one of Silicon Valley’s most talked-about AI and robotics startups.
But the moment you step inside, everything changes.
Instead of a front desk or welcome area, you walk straight into controlled chaos. The space feels raw and unfinished, like a large concrete box. Long wooden tables are spread everywhere. Some are used for lunch, covered with snack boxes, sauces, and random food items. Others are packed with computer screens, robot parts, wires, and robotic arms slowly learning how to do everyday human tasks.
This is not a normal office. This is where the future of Artificial Intelligence is being tested.
Robots Learning Everyday Human Tasks
All around the room, robotic arms are busy trying to copy simple human actions. One robot is attempting to fold a pair of pants, but it keeps making mistakes. Another tries to turn a shirt inside out but fails again and again. A third robot seems more confident as it carefully peels a zucchini, dropping the vegetable skin into a separate container.
These robots are not perfect. In fact, they fail often. But every failure is valuable data.
According to co-founder Sergey Levine, these machines are learning the same way humans do — by trying, failing, and trying again. Levine, who also teaches at UC Berkeley, explains it simply:
“Think of it like ChatGPT, but for robots.”
That one sentence captures the company’s big idea.
How Physical Intelligence Trains Robot Brains
The robots you see are part of a learning loop.
- Robots perform tasks in labs, homes, kitchens, and warehouses
- Every movement is recorded as data
- That data trains a powerful robotic AI model
- The improved model is tested again on real robots
This cycle never stops. Each task, whether folding clothes or peeling vegetables, helps the robot understand how the physical world works. If a robot learns how to peel a zucchini, it may later understand how to peel an apple or potato — even if it has never seen one before.
That ability to generalize tasks is the company’s real goal.
A Kitchen Built for Robots, Not Humans
One surprising part of the office is a fully equipped kitchen. There is even a professional espresso machine. At first glance, it looks like a perk for employees.
But it isn’t.
The kitchen exists only for robots.
Robots are trained here to handle cups, machines, liquids, and heat. If a robot successfully makes a latte, that’s not a bonus — it’s valuable training data. Every spill, mistake, and success helps improve the AI model.
In this startup, even coffee is Breaking News for engineers.
Cheap Hardware, Powerful Intelligence
Interestingly, the robotic arms used here are not expensive. Each one costs around $3,500. If made in-house, the cost could drop below $1,000. A few years ago, robots at this price level could barely do anything useful.
But Physical Intelligence believes smart software matters more than fancy machines.
Good intelligence, they say, can make simple hardware perform advanced tasks. This belief separates them from many traditional robotics companies.
Meet the Founder Behind the Vision
Another key figure at the company is Lachy Groom, a young entrepreneur from Australia. He sold his first company at just 13 years old and later worked at Stripe before becoming an early investor in famous startups like Figma and Notion.
After years of investing, Groom wanted to build something meaningful himself. When he met the researchers behind Physical Intelligence, he knew this was the opportunity he had been waiting for.
Today, the company has raised over $1 billion, making it one of the most well-funded robotics startups in Startup News.
No Rush to Make Money
One unusual thing about Physical Intelligence is that it does not promise investors quick profits. There is no clear timeline for selling products.
Most of the money is spent on computing power to train AI models. Groom openly admits that the company could raise even more money if needed.
This long-term approach is rare, but investors seem willing to wait.
Competing Visions in the Robot Race
Physical Intelligence is not alone. Other startups, like Skild AI, are also racing to build general robot intelligence. Some companies focus on selling products quickly, while others focus on deep research.
Skild AI has already launched commercial robots and claims strong revenue. Physical Intelligence, on the other hand, believes patience will lead to better results.
This debate — research first or sales first — is shaping the future of Silicon Valley robotics.
