Skild Brain

What it is

Skild Brain is the robot foundation model built by Skild AI. The company describes it as a general-purpose brain for machines rather than something built for one application or a constrained environment ^[2]. The distinction matters: a conventional robot controller is written or trained for a particular task in a particular setting, and breaks when either changes. A foundation model is instead trained on a very wide range of bodies and situations first, and adapted to specific work afterwards — the same pattern used for language models, which are trained on general text before being tuned for a particular job.

How it was introduced

Skild AI set out the idea in a post published on 29 July 2025, titled Building the general-purpose robotic brain ^[1]. In a second post on 24 September 2025, The case for an omni-bodied robot brain, the company argued for a model not tied to a single body plan, and said it had created a universe containing 100,000 different robots and trained its AI to control all of them ^[1].

Training across so many bodies is the technical core of the argument. A robot's control policy is normally shaped by the specific machine it runs on — the number of joints, the length of a limb, the kind of sensor it carries. If a model is exposed to a hundred thousand different configurations during training, the hope is that it learns something closer to the general problem of moving and acting, and less about any one chassis.

What the model does

Skild AI says the model demonstrated generalization and emergent capabilities across manipulation, locomotion and navigation on diverse robots ^[2]. In other words, skills that were not the target of any single training run appear once the model has seen enough variety — a common claim for large-scale pretrained models, and the main reason a single system might be reused rather than rebuilt.

Two settings are named specifically. The first is quadrupeds operating in adverse physical conditions ^[2], where the machine has to keep its footing on difficult ground. The second is vision-based humanoids carrying out dexterous manipulation for household and industrial tasks ^[2], where cameras rather than specialized sensing are the main input and hands have to handle objects with care.

The company also says the model can be fitted to existing robot hardware rather than requiring purpose-built machines ^[2]. That is a practical point as much as a technical one: robot makers already have hardware in the field, and a brain that can be installed on it is useful in a way that a brain needing new machines is not.

Where it sits

The commercial side developed alongside the research posts. In March 2026, Skild AI said it was partnering with ABB Robotics and Universal Robots to deploy its omni-bodied Skild Brain across industries ^[1]. The company also said it crossed US$100 million in annual recurring revenue ten months after its first commercial deployment ^[1] — annual recurring revenue being the yearly value of subscription-style contracts, a measure of how much usage has been sold rather than of profit.

Taken together, the claims describe a bet on generality: one model, many bodies, deployed on hardware that already exists. Whether that bet holds depends on how well the generalization reported by the company survives contact with messy real workplaces, and that is what deployments of this kind are meant to test.