skildAI

S1 is Skild AI's new robotics foundation model that learns a task from a single video demonstration and executes it with no fine-tuning or post-training . As shown in the launch, an operator records one clip of a task like making pancakes, potting a plant, or assembling a packed hygiene kit, drops it into the model's context window, and the robot performs the same steps in real time. Skild says S1 can follow horizons of up to ten minutes and generalize when scenes are rearranged, when objects are swapped, or when the human demonstration itself was imperfect. The pitch aimed at operators is about the cost of adding new skills. Skild claims that matching S1's one-shot accuracy with current vision-language-action models would take 50 to 100 hours of teleoperated data collection per task, which is the wall that has kept robots inside a small number of well-resourced deployments. By moving task specification from data pipelines to a video prompt, S1 is positioned for smaller sites like grocery stores, hospitals, and independent businesses that cannot run their own robotics data programs. Skild AI was founded in 2023 by Deepak Pathak and Abhinav Gupta, who left Carnegie Mellon professorships to build the company , and it is pursuing a single general-purpose model across robot embodiments rather than a per-robot stack. S1 is the first public release under that thesis and the reference point investors and operators will use to judge whether in-context learning transfers from language models to physical manipulation at useful reliability.