Skild AI S1 Learns New Robot Tasks From One Video With NVIDIA AI
Skild AI's S1 learns unseen robot tasks from one video, achieving 66% step success as NVIDIA and Foxconn deploy it for Blackwell assembly in factories.
Summary
Skild AI launched S1 in the week before September 10, 2026. The robot foundation model uses in context learning to interpret one operator video and execute unseen tasks without changing its weights or receiving task specific post training. It maps demonstrated intent, objects and sequences into actions, including work outside its pretraining data, while adjusting to moved objects, recovering from errors and recombining skills. S1 handles tasks lasting up to 10 minutes, including plant potting, pancake making, pour over coffee and kit assembly. A potting test moved from recording to autonomous hardware execution in 11 minutes.
On new multistep tasks, S1 achieved about 66% success per step versus 9% for a similar AI system, a more than sevenfold improvement. Skild estimates one short video provides the value of roughly 380 hands on examples that could take 50 to 100 hours to collect manually. The company reached a $100 million annual revenue run rate 10 months after its first commercial deployment and formed more than 60 partnerships across manufacturing, logistics, inspection, security, food preparation and other uses. Cofounder and CEO Deepak Pathak describes the shift as learning from experience instead of preprogramming.
Skild, NVIDIA and Foxconn are deploying Skild Brain on dual arm manipulators for NVIDIA Blackwell assembly. One workflow installs a busbar and limit block, fastens 16 screws and adapts to disturbances. NVIDIA infrastructure trains the shared brain with simulation, human video, teleoperation and permitted deployment data. Cosmos and Cosmos Curator process data, Omniverse and Isaac Sim create virtual environments, and Isaac Lab with Newton models forces, contact, collision and pressure. Nsight identifies training bottlenecks, while TensorRT accelerates inference. New GPU accelerated touch, grip and manipulation solvers will soon join Newton for all developers.
Positives
- S1 learns an unseen task from one video without weight updates, task specific post training or a new dataset.
- S1 achieved about 66% success per step, compared with 9% for a similar AI system.
- One short video can equal roughly 380 hands on examples that otherwise require 50 to 100 collection hours.
- Skild reached a $100 million annual revenue run rate within 10 months of its first commercial deployment.
- More than 60 partnerships span manufacturing, logistics, inspection, security and food preparation.
- Skild Brain is already handling a 16 screw NVIDIA Blackwell assembly workflow with Foxconn.
Risks & concerns
- S1 still fails about one third of steps in Skild's tests, a significant limitation for long, multistep industrial work.
- The 9% comparison baseline is identified only as a similar AI system, leaving its design and capabilities unclear.
- Commercial deployment experience can inform the broader model only when customer agreements permit its use.
- The GPU accelerated Newton solvers have no specific release date beyond being available soon.