Skild S1 Now Assembles Nvidia Blackwell Systems From One Demo Video
Skild AI’s in context robot model is running on dual arm lines at an Nvidia factory, fastening screws and installing busbars after a single video prompt.
What changed
Skild AI’s S1 robot foundation model learns unseen multistep manipulation tasks from a single video demonstration without weight updates, using in context learning. Nvidia’s 10 September 2026 blog confirmed Skild built S1 on Nvidia Cosmos, Isaac Lab, Isaac Sim, and Omniverse, and reported tasks up to ten minutes long with 66% multistep success versus 9% for a language prompted vision language action baseline after 100,000 hours of pretraining.
Coverage on 22 September 2026 highlighted a live deployment with Foxconn at an Nvidia factory in Houston, where dual arm robots assemble Blackwell GPU systems. The workflow includes fastening sixteen screws, installing a busbar, and recovering from physical disturbances without task specific retraining.
Skild said it reached a $100 million annual revenue run rate ten months after its first commercial deployment and maintains more than sixty partner deployments across manufacturing, logistics, and inspection.
Why it matters
Most warehouse automation still requires weeks of engineering per SKU change. A model that adapts from one demonstration shifts capital planning from fixed programming budgets to verification and safety review budgets. For semiconductor and electronics manufacturers facing rapid product refreshes, that changes ROI math on humanoid and dual arm lines.
Who is affected
Manufacturing operations leaders at electronics and automotive plants; robotics integrators competing with foundation model vendors; Nvidia ecosystem partners standardizing on Isaac and Cosmos training pipelines.
What to do next
Run a scoped pilot on one high mix, low volume subassembly and compare changeover time against your current teach pendant or offline programming baseline.
What to watch
Skild’s promised broader customer rollout after limited industrial partners, and published failure rates on safety critical torque steps beyond screw counts.
Sources
- Primary. Skild AI, Introducing S1: In Context Learning for Robotics (August 2026). Defines training recipe and benchmark claims.
- Primary. Nvidia, Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video (10 September 2026). Confirms infrastructure stack, revenue run rate, and Blackwell assembly deployment.
- Secondary. TechTarget, Nvidia backed Skild AI teaches robots new tasks from a single video (22 September 2026). Corroborates factory deployment framing.