Skild AI S1 Model Learns 10 Minute Robot Tasks From One Video, Replacing 380 Examples

The new S1 foundation model enables robots to execute complex physical movements by observing a human demonstration rather than relying on language instructions. Skild AI's S1 can learn multi-step processes lasting up to 10 minutes—including making pour-over coffee, assembling kits, or flipping pancakes—from a single video prompt without fine-tuning or post-training. In one performance test, a robot began autonomously potting a plant 11 minutes after the human demonstration was recorded.

S1 achieved a 66% success rate on unseen tasks in internal benchmarks, compared to 9% for a language-prompted Vision-Language-Action (VLA) model using the same 100K hours of pretraining. Skild estimates that one video prompt provides performance equivalent to roughly 380 task-specific post-training demonstrations. The model was trained on Nvidia AI infrastructure and is currently being deployed with a limited group of industrial partners ahead of a wider rollout in the coming months.