A four -legged robot that continues to crawl even after its four legs have been hacked with a chainsaw is the cloth of the nightmares for most people.
For Deepak Pathak, co -founder and CEO of the startup Skild AI, the dystopian feat of adaptation is an encouraging sign of a new kind of more general robotic intelligence.
“This is something that we call an omni committee brain,” said Pathak. His startup has developed the generalist algorithm of artificial intelligence to take up a key challenge with advanced robotics: “Any robot, all task, a brain. It is absurdly general. “
Many researchers believe that AI models used to control robots could feel a leap forward, similar to that which has produced language models and chatbots, if enough training data can be collected.
The existing methods for the formation of robotic AI models, such as the fact that algorithms learn to control a particular system through teleoperation or simulation, do not generate enough data, says Pathak.
Skild’s approach is to have a single algorithm Learn to control a large number of different physical robots through a wide range of tasks. Over time, this produces a model that the company calls Skild Brain, with a more general ability to adapt to different physical forms, including those that it has never seen before. The researchers created a smaller version of the model, called Looformer, for an academic article describing his approach.
The model is also designed to adapt quickly to a new situation, such as the missing leg or the new treacherous field, determining how to apply what he learned to his new situation. Pathak compares the approach to the way in which important language models can take particularly difficult problems by breaking it down and strengthening its deliberations in its own context window – an approach known as the learning in the context.
Other companies, including the Toyota Research Institute and a rival startup called physical intelligence, also run to develop more generally capable robot models. However, the skin is unusual in the way it builds models that are generalized through many types of equipment.
In an experience, the skin team has led to its algorithm to control a large number of walking robots in different shapes. When the algorithm was then executed on real two and four -legged robots – systems not included in the training data – it was able to control their movements and make them walk.
At one point, the team noted that a four-legged robot running the company’s omni-coded brain would adapt quickly when it is placed on its rear legs. Because it smells the ground under its hind legs, the algorithm exploits the robot dog as if it were a humanoid, making it walk on its posterior legs.
Generalist algorithm could also adapt extreme changes to the shape of a robot – when, for example, its legs were linked, cut or modified to become longer. The team also tried to deactivate two of the motors on a quadruped robot with wheels as well as legs. The robot was able to adapt by balancing two wheels like an unstable bike.
The skin tests the same approach for manipulation of robots. It has led to the brain of the skin on a range of simulated robotic weapons and found that the resulting model could control unknown equipment and adapt to sudden changes in its environment as a reduction in lighting. The startup is already working with certain companies that use robotic weapons, says Pathak. In 2024, the company raised $ 300 million in a round that estimated the company to $ 1.5 billion.
Pathak says that the results may seem frightening for some, but for him, they show the sparks of a kind of physical superintendent for robots. “It’s so exciting for me personally, guy,” he says.
What do you think of the Multiple Skild robot brain? Send an e-mail to ailab@wired.com to let me know.
This is an edition of Will Knight Ai Lab Newsletter. Read the previous newsletters here.
