Production lines and warehouses do not stay fixed. Tasks change, layouts shift, new products arrive, and most robots cannot keep up without significant reprogramming.
Skild AI's new S1 robot foundation model targets exactly that problem. It is designed to learn long-horizon tasks it has never seen before from a single video demonstration.
No retraining
The method does not rely on updating weights or on task-specific post-training. An operator records a video of the desired task and hands it to the model as a prompt; the model interprets the demonstrated intent, the objects and the sequence, then maps them into actions for the robot in front of it. This is called in-context learning.
S1 is said to perform unfamiliar tasks lasting up to ten minutes: plant potting, pancake making, pour-over coffee brewing, kit assembly. These span dozens of manipulation steps and require the robot to compose skills in sequences it has never performed.
The difference from the usual approach shows up here. Most industrial robots are built for fixed jobs; every new product, process or layout demands more data, retraining and validation. S1 aims to skip that cycle entirely.
The numbers
The measurements the company shared make the difference concrete.
| Measure | Value |
|---|---|
| Per-step success on a new multistep task | about 66% |
| The same measure for a similar AI system | 9% |
| What one short video is worth | roughly 380 hands-on examples |
| From demonstration to autonomous execution | 11 minutes in one test |
That last row may be the most striking. In a plant-potting test the team went from recording the demonstration to autonomous execution on hardware in eleven minutes.
The model can also adjust when objects move, recover from errors and combine skills in sequences that were never explicitly programmed. That is the capability that actually matters in the field; working in a lab with everything in its place is not enough.
NVIDIA's role
Skild built S1 and conducted the research on NVIDIA AI infrastructure. The collaboration spans synthetic data generation, model training, simulation and real-world deployment.
In the words of Skild AI cofounder and CEO Deepak Pathak, "learning by experience, and not preprogramming, is the step change that has happened in robotics". Pathak says NVIDIA Isaac Lab and Cosmos technologies help produce the scalable, diverse experience robots need to learn across many scenarios and embodiments.
The commercial side
The launch came as the company reached a 100 million dollar annual revenue run rate ten months after its first commercial deployment. In that same period it built more than 60 deployment partnerships.
Those partnerships span manufacturing, logistics, inspection, security and food preparation. So what is being described is not a laboratory demo; the model is meant to work in the field, in changeable environments. Which raises the real question: where the success rate lands when the demonstration video is poorer, or the environment differs from the expected one, has not yet been independently measured.