Skild AI Trains Robots to Do New Tasks From a Single Video

Pittsburgh-based startup Skild AI has introduced its S1 robot foundation model, also known as the Skild Brain. Built on Nvidia's AI infrastructure, this omni-bodied system allows robots to learn previously unseen, multistep tasks from a single video demonstration. The technology is already being put to work at an Nvidia factory in Houston. There, Skild AI, Nvidia, and manufacturing giant Foxconn are utilizing the model on dual-arm robots to assemble Nvidia Blackwell GPU systems.

The S1 model is designed to adapt when objects move, recover from errors, and combine skills in sequences without explicit programming. In one test, only 11 minutes elapsed between recording a video of a plant-potting demonstration and the robot autonomously performing the task. To achieve this speed, Skild AI leverages Nvidia's Cosmos, Isaac Lab, Isaac Sim, and Omniverse technologies for simulation, training, data generation, and deployment.

This rapid training capability addresses a major bottleneck in physical AI: the high cost and slow speed of collecting real-world physical data. Skild AI's commercial traction highlights the demand for this adaptability, with the startup reaching a $100 million annual revenue run rate this month, just 10 months after its first commercial deployment.

For robotics practitioners and manufacturers, this development signals a shift toward shared foundation models that can be reused across different machine types, including robotic arms, quadrupeds, and humanoids. Instead of programming robots for highly specific, rigid workflows, operators can quickly teach machines new tasks as products and processes change. This drastically reduces the development effort and integration time required to deploy adaptable automation on the factory floor.