Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video

Skild AI is pushing robots toward something closer to imitation than programming. Its new S1 foundation model, launched last week, takes a single video demonstration and turns it into action, without retraining the model or doing task-specific post-training. The pitch is simple: show the robot what to do once, then let it figure out the sequence, the objects and the intent on its own.

That matters because the real world keeps changing. Factory floors get rearranged. Warehouses get new products. Production lines don’t sit still just to make a robot’s life easier. Skild says S1 is built for that messier reality, where a task can run for up to 10 minutes and span dozens of steps, from plant potting to pancake making, pour-over coffee brewing and kit assembly.

The company says it can move from a recorded demo to autonomous execution on hardware in 11 minutes in one plant-potting test. It also says the model can recover from errors, handle moved objects and string together skills that weren’t explicitly programmed. On Skild’s own tests of new multistep tasks, S1 succeeded about 66% of the time at each step, versus 9% for a similar AI system. Skild also estimates that one short video example can be as useful as about 380 hands-on training examples.

NVIDIA is all over the stack here. Skild built S1 and ran the research on NVIDIA AI infrastructure, and the companies say they are working together across synthetic data generation, model training, simulation and deployment. NVIDIA’s Cosmos models help turn video into structured descriptions and diversify data. Isaac Lab, Omniverse, Isaac Sim and TensorRT cover the rest, from simulation to inference. The point is to keep robot learning tied together instead of treating data, training and deployment like separate planets.

The commercial part is moving too. Skild says it has reached a $100 million annual revenue run rate 10 months after its first commercial deployment, and now has more than 60 deployment partnerships across manufacturing, logistics, inspection, security and food preparation. It is also working with NVIDIA and Foxconn on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems, including a workflow that installs a busbar and limit block, fastens 16 screws and adapts to disturbances. That is the sort of slog robots have always struggled with. If video can teach it, the old factory playbook starts looking very expensive.