Skild AI unveils S1, which performs tasks lasting up to 10 minutes from video demos

Skild AI announced its robot foundation model, S1, on August 18, 2026. S1 is designed to turn a video demonstration of a task into robot actions, and the company says it demonstrated in-context learning on tasks it had not been trained on, including long-duration tasks lasting up to 10 minutes. The announcement emphasized that the model operates without additional training or fine-tuning after pretraining.

Skild AI’s announcement is an attempt to shift robots from “machines retrained for each task” toward “machines that assemble actions from video context.” The company’s claim that S1 can handle long-duration tasks of up to 10 minutes could point to broader applications than operations demonstrated with short clips, but reproducibility and safety still need to be verified.

The novelty of the S1 announcement is its effort to extend in-context learning in robotics from “short, familiar tasks” to “long-duration tasks of up to 10 minutes, including tasks the model has not been trained on.” Skild AI has described data collection and additional training for each new task as generally necessary in robotics; the latest claim raises the question of how far that assumption can be challenged. However, the announcement centers not on performance in general, but on an input design that translates video demonstrations directly into actions. In other words, its central idea is to bring the LLM-style approach of giving instructions in language to robotics learning centered on visual demonstrations. As shown by [Related coverage from this newspaper: Vention opens Physical AI Lab in Montreal to research robot manipulation for manufacturing](/news/vention-montoriooru-ni-physical-ai-lab-wo-kaisetsu-seizou-mu-04a96d) and LG Electronics announces plan to commercialize integrated actuators for humanoid robots at IFA 2026, the focus of physical AI is shifting from foundation models alone to the components, data and operations needed for real-world use. Skild AI’s S1 appears to be positioned within this trend as a test of how much diversity in real-world settings a model can absorb. At the same time, the materials released provide limited information about the scale of the training data, the number of physical robots used, the modalities covered, and the success rate on long-duration tasks. The next issues to examine are what kinds of tasks can be handled with a single video demonstration, how reproducible performance is on tasks the model has not been trained on, and how safety and intervention in the event of failure are designed when additional training is omitted.

The company said S1 can handle not only short tasks but also tasks it has not previously encountered during pretraining. Its evaluation distinguishes between tasks seen during training and those that were not, as well as between short tasks lasting about 5–25 seconds and long-duration tasks lasting around 10 minutes. The announcement says S1 was built on NVIDIA AI infrastructure and trained on diverse robot data covering different scenes, viewpoints and embodiments. The company also said it released an in-context learner for mobile robots a year ago, and that S1 extends this work to manipulation.

For Japanese companies, not only robot hardware but also video-based learning and methods for accumulating real-world robot data could become competitive factors. Companies with ways to connect to existing equipment and design data collection in manufacturing and logistics settings may be better positioned to identify practical uses for foundation models like these.