Skild AI’s universal robot brain goes starfish, adapts even after losing limbs

Skild AI has developed an “omni-bodied brain” that allows a single AI model to control multiple types of robots and adapt to extreme physical damage, including continuing to operate after losing limbs. The breakthrough represents a significant step toward more generalized robotic intelligence that could work across any hardware platform, addressing a key limitation in current robotics where AI models are typically designed for specific robot types.

What you should know: The AI system can control unfamiliar robotic hardware and adapt to severe physical modifications without additional training.

• When a four-legged robot’s limbs were cut off with a chainsaw, the AI algorithm quickly adapted to continue functioning with the remaining hardware.
• The model generalizes across different robot body types, successfully controlling both two- and four-legged robots that weren’t included in its original training data.
• In one test, a quadruped robot placed on its hind legs automatically began walking upright like a humanoid after sensing ground contact beneath its rear limbs.

How it works: Skild’s approach trains a single algorithm across numerous different physical robots performing various tasks, creating what the company calls “Skild Brain.”

• The model uses in-context learning similar to large language models, breaking down complex problems and feeding solutions back into its processing loop.
• Training involves gathering data from a wide range of robot types rather than focusing on individual systems through teleoperation or simulation.
• The AI can adapt to modifications like tied-together legs, extended limbs, or deactivated motors by leveraging its broad training foundation.

In plain English: Think of traditional robot AI like a specialist doctor who only knows how to treat one type of patient. Skild’s approach is more like training a general practitioner who has seen thousands of different cases and can adapt their knowledge to handle new situations they’ve never encountered before.

Beyond walking robots: Skild is applying the same generalist approach to robot manipulation tasks.

• The company trained Skild Brain on various simulated robot arms and found it could control unfamiliar hardware while adapting to environmental changes like reduced lighting.
• Skild is already working with companies that use robot arms, according to CEO Deepak Pathak.

The big picture: Multiple companies are racing to develop more generally capable robot AI models, but Skild’s approach of generalizing across diverse hardware types sets it apart from competitors.

• Toyota Research Institute and startup Physical Intelligence are also developing generalist robot AI models.
• Pathak believes this represents the emergence of “physical superintelligence for robots,” comparing the potential breakthrough to the leap that produced modern language models and chatbots.

What they’re saying: “Any robot, any task, one brain. It is absurdly general,” Pathak explained about their omni-bodied brain concept.

• “It is so exciting to me personally, dude,” he added regarding the results, though acknowledging they might seem “creepy to some.”

Follow the money: Skild AI raised $300 million in 2024 at a $1.5 billion valuation, reflecting investor confidence in generalized robotics AI.