Skild AI has unveiled S1, a robot foundation model designed to adapt to constantly changing industrial environments. Unlike traditional robotic systems that require substantial reprogramming whenever a task or layout changes, S1 aims to generalize from minimal data: a single video demonstration is reportedly enough for a robot to learn a new, even lengthy, sequence of actions.
This approach addresses a practical problem on factory floors, in warehouses and along production lines, where tasks shift frequently and the arrival of new products often forces a complete overhaul of automated system configurations. By cutting down on manual reprogramming, Skild AI aims to make robots more adaptable to these operational changes.
The S1 model was built using NVIDIA's Physical AI stack, a set of tools and platforms meant for training robotic systems capable of interacting with the physical world. The collaboration reflects NVIDIA's broader strategy of positioning itself as an infrastructure provider for AI applied to robotics, extending beyond its established role in compute for generative AI.
While S1's precise generalization capabilities still need to be validated in real-world industrial deployments, the announcement fits into a wider trend: robot foundation models seeking to replicate, for physical manipulation, what large language models achieved for text — namely, the ability to learn quickly from very few examples.