Skild S1 Learns Multistep Assembly Tasks From One Video on Blackwell Lines
Skild AI and NVIDIA disclosed on 10 September 2026 that Skild Brain on dual arm manipulators is assembling NVIDIA Blackwell systems at Foxconn, learning multistep tasks including sixteen screw installs from a single demonstration video.
What changed
On 10 September 2026, NVIDIA published a collaboration brief with Skild AI describing the S1 robotic foundation model and live deployment on Foxconn factory floors. Skild Brain now runs on dual arm manipulators performing high precision assembly of NVIDIA Blackwell systems, including busbar and limit block installation with 16 screw sequences and recovery from disturbances.
Skild trains S1 on NVIDIA infrastructure spanning Cosmos synthetic data, Omniverse and Isaac Sim simulation, teleoperation, and deployment feedback, with TensorRT optimizing edge inference latency.
Why it matters
Manufacturing leaders have heard years of pick and place demos that fail when lighting, torque, or sequence order shifts. Skild's claimed workflow adapts across multistep tasks learned from one video, which matters for fabs and contract manufacturers that cannot afford bespoke programming per SKU.
Foxconn's Blackwell line is a high stakes reference: if the system holds torque and sequence accuracy on NVIDIA's own production hardware, other electronics assemblers will treat Skild as a credible generalization layer rather than a lab curiosity.
Who is affected
Plant managers at electronics and automotive tier ones, integrators pairing Universal Robots or ABB arms with AI brains, and NVIDIA ecosystem partners selling Isaac and Omniverse should expect customer RFPs asking for video to task generalization evidence on their SKUs.
Competing robot foundation model vendors must respond with deployed screw driving or insertion metrics, not simulation only reels.
What to do next
If you operate high mix assembly cells, request a pilot scope that mirrors Skild's disclosed Blackwell task: multistep fastening, disturbance injection, and cycle time after a single demonstration clip. Compare against your incumbent vision plus teach pendant baseline on identical throughput and rework rate.
What to watch
Whether Skild or Foxconn publish audited uptime, mean time between intervention, and yield data from the Blackwell cell through Q4 2026, and whether ABB or Universal Robots partnership deployments named at GTC show comparable task breadth.
Sources
- Primary. NVIDIA Blog, Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video (10 September 2026). S1 training stack, Foxconn Blackwell workflow, and 16 screw sequence detail.
- Secondary. NVIDIA Newsroom, NVIDIA and Global Robotics Leaders Take Physical AI to the Real World (2026). Skild partnerships with ABB, Universal Robots, and Foxconn context.