This AI-Driven Robot Remains Functional Despite Chainsaw Assaults

A robot with four legs that continues to move even after losing all of them to a chainsaw is a nightmare scenario for many.

For Deepak Pathak, cofounder and CEO of Skild AI, this unsettling display of adaptability signals a new form of robotic intelligence.

“We refer to this concept as an omni-bodied brain,” Pathak explains. His startup has designed a generalized artificial intelligence algorithm aimed at overcoming significant robotic challenges: “Any robot, any task, one brain. It’s incredibly broad.”

Many researchers believe that with sufficient training data, the AI models that direct robots could undergo a transformative advancement, akin to the development of language models and chatbots.

Pathak notes that traditional methods for training robotic AI, such as using teleoperation or simulations, fail to produce sufficient data.

Skild’s strategy is to develop a single algorithm capable of controlling various types of robots across numerous tasks. This approach, producing a model referred to as Skild Brain, enables adaptability to different physical forms, even those unseen before. For academic purposes, they created a smaller version named LocoFormer to illustrate their method.

The model is also engineered for rapid adaptation to new challenges, like a missing leg or dangerous terrain, figuring out how to leverage its existing knowledge in unfamiliar circumstances. Pathak likens this technique to how large language models tackle complex issues by breaking them down and recontextualizing their insights—known as in-context learning.

Other organizations, including the Toyota Research Institute and a competing startup called Physical Intelligence, are also striving to create broadly capable robotic AI models. Skild stands out for its focus on developing models that generalize across a diverse array of hardware.

In a recent experiment, the team instructed their algorithm to manage multiple walking robots with different designs. When tested on actual two- and four-legged robots—not included in the training set—it successfully directed their movements, enabling them to walk around.

At one point, the team discovered that a four-legged robot using the omni-bodied brain quickly adjusted when it was made to stand on its hind legs. By sensing the ground beneath it, the algorithm operated the robot as if it were a humanoid, allowing it to walk upright.

The generalist algorithm also has the capacity to adjust to drastic changes in a robot’s configuration—like when its legs are tied, removed, or lengthened. Additionally, the team tried disabling two motors on a quadruped robot equipped with wheels and legs, and it managed to balance on two wheels much like a wobbly bicycle.

Skild is applying the same methodology to robotic manipulation. They trained Skild Brain on various simulated robotic arms and found that the resulting model effectively controlled unfamiliar hardware and adapted to sudden changes in environmental conditions, like decreased lighting. Pathak mentions that the startup is already collaborating with companies utilizing robotic arms. In 2024, they secured $300 million in funding, bringing the company’s valuation to $1.5 billion.

Although Pathak acknowledges that the findings might seem unsettling to some, he views them as indications of the emergence of a unique form of physical superintelligence in robots. “It’s incredibly exciting for me, dude,” he shares.

What are your thoughts on Skild’s versatile robotic brain? Feel free to reach out via email at ailab@wired.com.

This is an edition of Will Knight’s AI Lab newsletter. Read previous newsletters here.