Skild AI unveils S1, a robot foundation model that picks up a task from one video

Skild AI released S1, its flagship "omni-bodied" robot foundation model, on 31 August 2026. It runs across quadrupeds, humanoids and static arms, learns new multi-step tasks in context from watching a single video, and holds jobs of up to ten minutes — repotting a plant, brewing coffee, cooking pancakes. It is trained on four stacked data sources: teleoperation, human video, simulation and data-capture glove recordings, which CEO Deepak Pathak argues is necessary because each one covers another’s weakness. Pathak says pancake-flipping emerged without an explicit training example, and that production deployments will be announced separately.

What it means One-video, in-context task acquisition is the demo the whole robot-foundation-model field is chasing — and no benchmark score or deployment count has been published against it yet.