10 Sep 2026
NVIDIA blog: Skild AI’s S1 learns long robot tasks from one video
NVIDIA published that Skild AI’s S1 robot foundation model can learn previously unseen long tasks from a single video demonstration, using the video as the prompt and without updating its weights.
NVIDIA documents a path where a robot foundation model takes one operator video as the prompt for multistep factory tasks, with a Foxconn/Blackwell assembly example. Capability numbers and the $100 million annual revenue run rate stay the companies’ claims.
NVIDIA published “Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video.” S1 is designed to learn previously unseen long-horizon tasks — jobs that take many steps — from a single video demonstration. The model reads demonstrated intent, objects, and sequence, then maps them to actions without updating weights or running task-specific post-training. That is in-context learning: the video is the prompt. That NVIDIA blog is the primary.
NVIDIA and Skild say S1 can perform unfamiliar tasks lasting up to about 10 minutes — plant potting, pancake making, pour-over coffee brewing, kit assembly — spanning dozens of manipulation steps and skill sequences the model had not previously performed.
Skild tests, as reported on NVIDIA’s page: on new multistep tasks, S1 succeeded about 66% of the time at each step versus 9% for a similar AI system — more than sevenfold. Skild estimates one short video can be as useful as roughly 380 hands-on training examples, which a person collecting manually could take 50–100 hours. Those figures are Skild’s.
NVIDIA says Skild built S1 on NVIDIA AI infrastructure. The collaboration spans synthetic data (Cosmos), Isaac Lab / Isaac Sim, Omniverse, and TensorRT. Newton physics and joint GPU-accelerated simulation solvers are mentioned as joint work, with solvers to be made available in Newton. Stack wiring is in Sources.
Deepak Pathak, Skild CEO, on NVIDIA’s page: “Learning by experience, and not preprogramming, is the step change that has happened in robotics.” He cited Isaac Lab and Cosmos for scalable, diverse experience across scenarios and embodiments.
The same post’s commercial framing: Skild reached a $100 million annual revenue run rate 10 months after its first commercial deployment, and has more than 60 deployment partnerships spanning manufacturing, logistics, inspection, security, food preparation, and other applications. Treat that as Skild’s claim via NVIDIA, not audited financials.
On a factory demo, Skild, NVIDIA, and Foxconn are deploying Skild Brain on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems. NVIDIA describes a demonstrated workflow: a robot installs a busbar and limit block, fastens 16 screws, and adapts to disturbances across a multistep task.