Skild AI unveils S1, a foundation model that learns ten-minute tasks from a single demo

Skild AI has launched S1, a foundation model that performs long-horizon manipulation tasks — exceeding ten minutes — after watching a single demonstration video, with no fine-tuning. The model operates through in-context learning: it takes a video as a prompt and outputs a sequence of robotic actions that completes the task in any environment and on any robot embodiment.

In practice, the company demonstrates coffee preparation, planting a pot, and frying pancakes — tasks that never appeared in the pre-training data. When the team searched the training corpus for pancake-flipping examples, they found none; S1 inferred the action from the single video. Unlike a language model that generates text, the output here is motion trajectories and motor control, and the model is designed to generalize to a robot body different from the one in the video.

On familiar tasks, Skild says S1 matches the performance of vision-language-action (VLA) models driven by language prompts. The advantage appears on novel tasks: as pre-training scale grows, the gap widens exponentially. The company claims that to reach the same accuracy S1 achieves from one demonstration, current VLA models would require 50 to 100 hours of additional data collection and training.

The model does not simply replay the demonstration. According to Skild, it exhibits common-sense understanding: it is robust to disturbances, improvises when errors occur — even if the human in the video did not correct them — and sometimes executes with higher precision than the human demonstrator. If the claims hold, this represents a paradigm shift: instead of gathering hundreds of demonstration hours for each new task, a single video suffices.

Last year Skild introduced Locoformer, which demonstrated in-context learning for mobility: controlling any robot and adapting to disturbances such as broken legs or locked wheels. S1 extends that same capability to general manipulation. The company calls the result an "omni-bodied brain"; if the robot can move, the brain will move it, even in an entirely new body.

S1 is already deployed with a limited set of industrial partners, with expansion to additional customers planned in the coming months. Skild describes the model as "the first light on a new path to general intelligence grounded in the physical world." A detailed technical blog post is expected later.