Skild AI Uses Nvidia Tech to Teach Robots New Tasks From a Single Video
Pittsburgh startup Skild AI has deployed its S1 foundation model, which allows robots to master complex physical workflows from a single video demonstration, while reaching a $100 million revenue run rate.
Teaching Robots Complex Tasks From Single-Video Demos
Pittsburgh-based robotics startup Skild AI has introduced S1, a physical AI foundation model designed to learn complex, multi-step tasks from a single video demonstration. Backed by Nvidia, the company is targeting general-purpose robotics by drastically cutting the time and programming required to transition automated systems from research labs to production environments.
The commercial rollout comes as Skild AI hit a $100 million annual revenue run rate, achieving the financial milestone just 10 months after its initial commercial deployment.
Houston Factory Deployment and the S1 Architecture
Skild's technology is already running in production. At an Nvidia manufacturing facility in Houston, Skild, Nvidia, and electronics manufacturer Foxconn are using dual-arm robots powered by the Skild Brain omni-bodied foundation model to assemble Nvidia Blackwell GPU systems.
The S1 model was developed on Nvidia’s software and computing infrastructure, leveraging Cosmos, Isaac Sim, Isaac Lab, and Omniverse to handle data generation, simulation, model training, and deployment. Once deployed, the system can autonomously adapt to moved items, recover from operational errors, and string together skills in sequences it was never explicitly programmed to follow.
In one trial run, Skild demonstrated that just 11 minutes passed between recording a video of someone potting a plant and the robot autonomously replicating the task.
"S1 shows how quickly an operator could teach a robot new work," said Amit Goel, director of product management for autonomous machines at Nvidia. "For manufacturers, the opportunity is to make automation more adaptable as products, components and processes change."
Solving Physical AI’s Data and Embodiment Challenges
Despite rapid progress in foundation models, physical AI remains constrained by the difficulty of gathering real-world training data. Unlike digital AI systems that ingest text or code from the internet, robotic systems must master variations in lighting, object placement, tactile feedback, and unpredictable edge cases.
"Collecting that experience on real robots is slow and expensive," Goel noted, pointing to simulation platforms like Isaac Lab and Omniverse, alongside synthetic data generation via Cosmos, as critical workarounds. The ongoing challenge lies in ensuring that simulated testing accurately translates into reliable performance on factory floors.
Looking ahead, developers are aiming for shared foundation models capable of powering diverse robot forms. Skild AI has run its models across robotic arms, quadruped robots, and humanoid platforms. According to Goel, establishing reusable machine intelligence across varied mechanical bodies and factory tasks represents the next critical milestone for physical automation.
