Article illustration 1

When researchers at Skild AI hacked the legs off a quadruped robot with a chainsaw, the machine didn't collapse—it adapted its movement strategy in real-time and kept crawling. This startling demonstration showcases what CEO Deepak Pathak calls an "omni-bodied brain": a single AI model capable of controlling any robotic form, even those it's never encountered before, while dynamically adjusting to physical trauma or environmental chaos.

Beyond Fragile Specialization

Traditional robotics AI relies on painstakingly tailored models for specific hardware configurations, making them brittle when facing unexpected changes. "Existing methods like teleoperation or simulation don't generate enough diverse data," Pathak explains. Skild's breakthrough comes from training one algorithm across thousands of different simulated robot bodies—varying limb counts, joint configurations, and weights—while subjecting them to aggressive randomized challenges. The resulting model, dubbed Skild Brain, develops a generalized understanding of physical embodiment that transfers to real-world hardware.

"Any robot, any task, one brain. It is absurdly general," says Pathak, comparing the system's adaptability to large language models that use in-context learning to break down novel problems.

Extreme Adaptation in Action

In peer-reviewed research (using a smaller variant called LocoFormer), Skild's AI demonstrated remarkable resilience:
- A quadruped stood upright and walked on hind legs when tipped backward
- Robots with legs tied together, extended, or amputated improvised new gaits
- A quadruped with deactivated wheels transformed into a wobbly two-wheeled balancer

The system continuously learns from failures—if a robot falls, the AI updates its strategy for subsequent attempts without human intervention. This online adaptation stems from reinforcement learning at massive scale, where the model "rebuilds representations" of its body after disturbances.

The Race for Robotic Generalization

Skild isn't alone in pursuing versatile robot intelligence; Toyota Research Institute and startup Physical Intelligence share similar ambitions. But Skild stands out by generalizing across radically different morphologies. The approach extends beyond locomotion: the same architecture controls unfamiliar robot arms and adapts to environmental shifts like sudden darkness.

With $300 million in recent funding at a $1.5 billion valuation, Skild is already deploying early versions with industrial partners. Pathak sees this as the foundation for "physical superintelligence"—though he acknowledges the dystopian optics. "It is so exciting to me personally, dude," he admits.

As boundaries between simulation and reality blur, robots that withstand extreme damage could transform fields from disaster response to space exploration. Yet this resilience also raises questions: How do we ensure such adaptable systems remain predictable? The chainsaw test proves what's possible; the next challenge is shaping how these minds evolve.

Source: WIRED