TNS
Picture Credits: iStock
Skild AI just raised nearly $1.4 billion, taking its valuation past $14 billion, to build one “general brain” that can run any robot, on any task.
East Liberty’s Skild AI, making one ‘brain’ for every robot, now valued at $14 billion
Author: Chloe Jad
Any robot. Any task. One brain.
That’s the driving mantra of Skild AI, the East Liberty robotics company working to create a general brain for robots. And, on Wednesday, Skild AI announced it raised nearly $1.4 billion to bring its valuation to over $14 billion.
“The Skild Brain can control robots it has never trained on, adapting in real time to extreme changes in form or environments,” said Deepak Pathak, CEO and co-founder of Skild AI. “The model is forced to adapt rather than memorize — much like intelligence in nature.”
Leading investors in the latest round of funding included Softbank Group, Jeff Bezos’ Bezos Expeditions, NVIDIA’s venture capital arm NVentures, Macquarie Capital, Disruptive and 1789 Capital.
“We believe that a unified, omni-bodied brain is the fastest way to establish a continuous data flywheel where the model gets better with every single deployment, no matter what the hardware or task,” Mr. Pathak said.
While traditional software models are tailored to specific robot designs, “the Skild Brain is omni-bodied and can control any robot without prior knowledge of their exact body form, including quadrupeds, humanoids, tabletop arms, and mobile manipulators,” the announcement states.
That enables robots with a Skild Brain to do things like load a dishwasher, make an egg or navigate stairs and slippery terrain — “If there is a machine that can move, the omni-bodied Skild Brain will eventually be able to operate it.”
The smartphones of the next generation
What Skild — and most other AI and robotics ventures — care about most is data.
“Data has been the main driver of success in AI,” Mr. Pathak said at the second annual AI Horizons summit in September. “Such data is completely unavailable for robotics.”
ChatGPT’s level of generality, being able to ask for virtually any task in any context, is what the robotics innovators are aiming for with physical AI (or AI embodied in real-world entities) to achieve. The problem is, while general AI models train on trillions of data points, robotics have no equivalent.
Robots need to be deployed in the real world to gather data.
Because they must interact with the environment around them, robots not only train on language, just like large language models (LLMs), but also on vision, touch and reactivity to real-world stimuli. Hardware is an expensive and time-intensive endeavor, while software is more nimble and marketable. So the “recipe” is not the same for scaling robots as it is for ChatGPT, Mr. Pathak said in September.
While LLMs might train on 100 trillion examples, Mr. Pathak said it would take centuries to gather the same data set for robotics, even with billions of employees.
Skild AI is working on training its robots based on videos of human behavior and as many sources of training data it can get its hands on. Even so, Mr. Pathak said we need robots to do superhuman things, so training around a human-shaped robot is not always the best option, and it’s far from the only form. One general brain would be able to adapt to any robot, any task.
In a video of its progress, Skild AI’s suite of robots can be seen climbing flights of stairs, leaping over obstacles in parkour form, placing wireless earbuds back into its case, making coffee with a Keurig, sorting objects and even staying upright after an employee delivered a flying kick to the 6-foot humanoid robot’s chest.
These robots are not like the static robots in factories that aren’t aware of the world around them, and only perform predetermined actions and motions — they are reacting in real time without every move controlled by a human. The robot sees stairs and lifts its foot. The robot feels a tug or a push and keeps its balance.
“This is not just for show, these robots have to make useful cases for themselves,” Mr. Pathak said at AI Horizons.
They are expensive and time-intensive, and don’t make money back the way software does, so their real-world applications must provide value, he said.
And while much research still has to be done, robots will become the smartphones of the next generation, Mr. Pathak said, with each household having two robots — one for work, and one for home.
Learning like babies do
For the roboticists at Skild AI, humans are really just biological robots — with muscles and joints instead of metal and wires, said Mr. Pathak in September, but the same principles drive both bodies and their movements.
“The Skild Brain is not only omni-bodied but also demonstrates the ability to adapt to unpredictable scenarios, such as loss of limbs, jammed wheels, and increased payload, or even an entirely new body, without retraining or fine-tuning,” Wednesday’s announcement explained.
In an August interview with the Post-Gazette, Mr. Pathak said he and co-founder Abhinav Gupta wanted their robots to learn in an “active manner… exactly how babies learn.”
“We throw things, we put things in our mouth, and we basically start to learn about the world. Intelligence is born out of these kinds of physical experiments,” he said.
And while the race to conquer and deploy AI in humanoid form is led by the Chinese company AgiBot, Mr. Pathak noted that what Skild is doing with its omni-bodied brain is novel.
“What we are building is one of the most unique companies out there,” Mr. Pathak said. “If you look at today’s AI era, everyone out there is kind of building what we call digital intelligence, training mainly on digital data like language … while what we have been arguing is that general intelligence can only be built through grounding in the physical world.”
Credits: TCA, LLC.