Generalist AI’s Gen-1 model aims to “teach robots common sense of physics.”
Katie Collins Main writer
Katie is a UK-based journalist and features editor. Officially, she is CNET’s European correspondent, covering technology policy and Big Tech in the EU and UK. Unofficially, she is Taylor Swift’s correspondent on CNET. You can also find his writings on technology for good, ethics and human rights, the climate crisis, robots, travel and digital culture. She was once described as a “living synthesizer” by London’s Evening Standard for having a microchip injected into her hand.
In 2026, we see robots advancing by leaps and bounds with vastly improved dexterity, the kind of progress long needed in the quest for truly useful household helpers. Now, a new AI model has arrived to power robots through activities including folding laundrybuild boxes, repair other robots and even fill wallets with flimsy paper money.
Earlier this month, California-based Generalist AI launched Gen-1, a new Physical AI model that makes robots capable of performing all of these tasks (and more) successfully. It’s a big step forward in terms of robots designed for the real world and based on intelligence born from the real world, Pete Florence, co-founder and CEO of Generalist AI, told me.
In most of the sample videos released by the company, the Gen-1 operates on a pair of robotic arms, but that’s not the only reason it’s designed for. “Gen-1 is designed to be the brain of any robot, meaning the same model can work on a humanoid, industrial arm or other robotic systems,” Florence said.
This has already proven to be a groundbreaking year for general-purpose humanoid robotswith companies including Boston Dynamics And Honor unveiling cutting-edge robots capable of strangely human movements. The robot market is expected to explode, with one estimate from Morgan Stanley predicting growth to reach a $5 trillion market by 2050. Forecasts predict that robots will arrive in manufacturing, retail, hospitality and healthcare environments before eventually landing in our homes. To achieve this, we need to see further progress in the field of AI.
Training robots to live alongside humans
In recent years we have seen major linguistic patterns emerge, such as ChatGPT, Gemini and Claudeevolve at lightning speed. The same is not true for the physical AI models needed to power robots, largely because of the lack of data on which to train these models. Robots – and especially humanoid robots – must learn to navigate a world built for humans, just as a human would.
Often, this data is collected from robots performing tasks while being teleoperated by humans, but not by Generation 1. Instead, the dataset used to train Generalist AI’s models was assembled by humans performing millions of different tasks using wearable technology.
“We built our own lightweight ‘data hands’ and distributed them globally to learn how people actually interact with objects, with all the subtle force feedback, tactile sensation, drags, corrections and recoveries that define human dexterity in the real world,” Florence said. “This type of data is essential for teaching robots physical common sense, intuitive understanding and the ability to adapt in real time rather than carrying out rigid instructions.”
Generalist AI released a series of videos showing the model running on robots repetitively performing a range of different tasks, with perhaps the most compelling being a robot taking money out of a wallet before inserting it back into the same pocket. This is a tricky task that many humans struggle with. It’s clearly not easy for the robot either, given the fragility of paper money and wallet fabric – and yet it accomplishes the task.
Another video shows a robot sorting socks by color, folding them into neat piles and counting the number of pairs using a touchscreen. Other tricky tasks the model can accomplish include unzipping and filling a pen case, stacking oranges in a neat pyramid, and plugging in an Ethernet cable.
These videos show the extent of Gen-1’s abilities, but what’s more impressive is the success rate with which he can accomplish certain tasks. General AI measured the model’s success rate against the previous version and found that Generation 1 could successfully handle a robot vacuum cleaner in 99% of cases (compared to 50% for generation 0), folding boxes in 99% of cases (compared to 81% for generation 0) and wrapping phones in 99% of cases (compared to 62% for generation 0).
Robots improvise
Most robots are programmed to complete a task in a specific and orderly manner. But what happens when a curveball is thrown? “The smallest changes in the environment can cause failures,” Florence said.
An important skill that robots need, and that humans naturally possess, is the ability to think on their feet. That’s why Gen-1 was designed with improvisation in mind to be able to come up with strategies to accomplish tasks. Florence gives me the example of a robot using both hands to reposition a misplaced part for an automotive task, even though it was only trained to use one.
“Until now, this type of creativity has been largely absent from robotics,” he said.
Significant work remains to be done to strengthen robots’ improvisational capabilities, but early progress shows glimpses of a positive impact on both reliability and speed, Florence says. “We are starting to see real progress and are excited to push the boundaries of embodied intelligence.”
After all, one day you may need a robot in your house that can fix all your other smaller robots.
