A Robot Watched One Video and Flipped a Pancake. Skild AI Says That Changes Everything.
Skild AI's new S1 model lets robots learn tasks from a single video clip, without weeks of retraining. The company has raised nearly $1.7 billion to make it happen.
Key points
- Skild AI launched S1, its flagship robot foundation model, in 2025, claiming robots can learn new tasks from a single video demonstration.
- The company has raised nearly $1.7 billion since its founding in 2023 to build what it calls a general-purpose robot brain.
- S1 uses in-context learning, a technique that lets a model follow instructions or examples given at the moment of use, with no extra retraining required.
- Skild CEO Deepak Pathak confirmed a pancake-flipping behaviour emerged on its own, without any flipping examples in the training data.
- S1 can run on multiple robot body types, including robotic arms, four-legged robots and humanoids.
Teaching a robot a new skill normally takes weeks. Engineers collect examples, retrain the software, test it, repeat. Skild AI wants to cut that entire process down to showing the robot a video.
The Pittsburgh-based startup this week revealed S1, its first full robot foundation model. A foundation model is a large AI system trained on vast amounts of varied data, the same kind of technology that sits underneath ChatGPT. Instead of building a narrow system for one job, the idea is to build one brain flexible enough to handle many.
How does learning from a single video actually work?
S1 uses in-context learning, meaning the robot is shown an example at the moment it needs to act, rather than going through a separate training process. You show it a video, and it tries to copy what it sees on real hardware.
"You just add a video of a human doing something in the prompt, also called the context of the model, and it can just follow it on the robot," Pathak told The Robot Report.
The tasks Skild is targeting are not quick. We are talking about jobs that take up to ten minutes, like repotting a plant, brewing coffee, or cooking pancakes from scratch.
The pancake example is striking. When the robot flipped a pancake successfully, Skild's team went back through millions of hours of training data looking for any flipping motion to explain it. They found none. The behaviour had emerged from the model watching how a spatula moves, not from a direct example.
What data does S1 train on?
Most robot AI companies rely on one main source of training data. Skild uses four.
| Source | Strength | Weakness |
|---|---|---|
| Teleoperation (human drives robot directly) | Highest quality, straight from hardware | Slow to collect, low variety |
| Human videos | Huge volume, very diverse | Hard to apply directly to robot bodies |
| Simulation (virtual environments) | Cheap and scalable | Gap between virtual and real world |
| Data-capture gloves | More scalable than teleoperation | Slightly less direct than hardware |
"The pros of one source compensates for the downside of another source," Pathak said.
S1 also runs across different robot body types: tabletop arms, four-legged machines and humanoid robots. Pathak says stronger humanoid performance is a priority for future work.
Is this the moment robots become as useful as ChatGPT?
Not yet, by Pathak's own account. Before ChatGPT, language AI needed to be retrained for every new task. ChatGPT changed that by letting users simply type what they needed. Pathak sees S1 as a similar shift for robots, but an early one.
"Is it completely ready to be rolled out to people's homes? Not quite. But this is the first sign of what we believe might come," he said.
Skild recently acquired Fetch Robotics, a company with experience getting robots onto real warehouse floors, to speed up practical deployments. Commercial results are expected in the coming weeks.
For now, S1 is a strong signal that robot training is getting faster and more flexible. Whether that translates to robots doing useful work in ordinary settings is the next question to watch.
