StorySoftware

Skild AI releases S1, a model that learns new tasks from a single video demonstration

Skild AISkild Brain

Skild AI released S1, a robot model that watches one video of a task and then performs it, with no retraining. Trained on 100,000 hours of data, it succeeded 66% of the time on tasks it had never seen, against 9% for a comparable model told what to do in words. Skild says one demo video replaces what would otherwise take 50 to 100 hours of remote-controlled robot practice to collect.

  • Examples: potting a plant, cooking pancakes, making pour-over coffee, assembling a kit; tasks ran up to 10 minutes.
  • In the plant-potting example, 11 minutes passed from demonstration to the robot doing it alone.
  • On familiar tasks, conventional fine-tuning still wins: 86% versus 66% for S1.
  • It struggles when the job requires a different plan from the one shown in the video; tests allowed human help to recover from failures.
Read the original · Blog

More on Skild Brain

Primer