Teaching a robot something new has always been the slow part. You either write the instructions in code, or you record hundreds of demonstrations and retrain the model. Days of work for one task.
Skild AI says it has skipped that step. On August 25 the company released S1, a robot foundation model that learns from a single video. Film yourself doing something — potting a plant, packing a box, sorting parts — hand the clip to the robot, and it copies the job. No retraining. No new code.
Here's the breakdown:
- One video instead of 380 examples. Skild says a single demonstration is worth roughly 380 traditional training examples.
- Ten-minute tasks. S1 handles jobs with dozens of steps, not just "pick up the cup."
- 66% vs 9%. On tasks it had never seen before, S1 succeeded about 66% of the time. The same kind of system, prompted with written instructions instead of video, managed 9%.
- 11 minutes from film to work. In Skild's plant-potting demo, that's all the time between recording the video and the robot running on its own.
Why it matters: video carries things words can't. How hard to grip. What angle. What "finished" looks like. It's the difference between reading a recipe and watching someone cook — and it means the person teaching the robot doesn't need to be an engineer.
The catch: there are no open weights, no public API and no peer-reviewed paper. The results come from four hand-picked tasks on Skild's own benchmark. Treat 66% as a direction, not a number you can bank on. Skild raised $1.4B at a valuation above $14B in January, so the incentive to demo well is very real.
Still, the direction is hard to miss: soon, teaching a machine may just mean letting it watch you.
