Skild AI's S1 robot model leans on NVIDIA Isaac Lab and Cosmos to learn from single-video demos

Skild AI, a robotics foundation-model startup, says its S1 model can pick up new physical tasks from a single video demonstration without retraining, and NVIDIA detailed the technology behind that capability on its blog September 10. The two companies, along with Foxconn, are now deploying the system on Foxconn's NVIDIA Blackwell assembly lines.

What's new

NVIDIA's post lays out the mechanics and the traction:

• S1 learns via in-context learning rather than fine-tuning, letting it pick up a new task from one demonstration and apply it across different robot bodies and environments.
• "NVIDIA Isaac Lab and NVIDIA Cosmos technologies help Skild create the scalable, diverse experience its robots need to learn across many scenarios and embodiments," according to NVIDIA.
• In tests cited by NVIDIA, S1 succeeded on 66% of task steps, compared with 9% for competing systems it was benchmarked against.
• Skild has more than 60 deployment partnerships across manufacturing, logistics, and inspection, and NVIDIA says the company has reached a $100 million annual revenue run rate.
• Skild, NVIDIA, and Foxconn are now running the technology on Foxconn's Blackwell chip-assembly lines — a production use case rather than a lab demo.

Skild cofounder and CEO Deepak Pathak framed the shift as a break from the old robotics playbook: "Learning by experience, and not preprogramming, is the step change that has happened in robotics."

Context

Skild has positioned itself as a builder of general-purpose "robot brains" that can transfer across hardware platforms rather than being hand-coded for a single robot arm or task. That only works if the model has seen enough varied physical experience to generalize — which is expensive and slow to collect on real hardware. NVIDIA's Isaac Lab (a simulation and reinforcement-learning framework) and Cosmos (a world-model/video-generation platform for synthetic physical data) are both aimed at solving that data problem by generating simulated experience instead of relying solely on real-world robot time.

The Foxconn deployment matters because it moves the collaboration from a research showcase into an actual manufacturing environment tied to NVIDIA's own hardware supply chain — Blackwell chip assembly. That is a notably concrete proving ground for a foundation-model robotics company still early in commercialization.

Why it matters

The claimed 66%-versus-9% step-success gap, if it holds up outside NVIDIA's own benchmarking, would be a meaningful signal that video-conditioned, simulation-trained policies are starting to close the gap between "robot demo" and "robot that reliably does the job on a factory floor." A $100 million revenue run rate and 60-plus deployment partnerships also suggest Skild has moved past pilot-stage interest into paying, repeat customers faster than most robotics-model startups.

For NVIDIA, the partnership is a proof point for its Isaac Lab and Cosmos stack as the default simulation layer for physical AI, reinforcing a strategy of selling not just the chips that run robots but the tools used to train them. Coupled with the Foxconn tie-in, it also signals NVIDIA's own hardware production is becoming a showcase customer for the physical-AI ecosystem it is trying to build — a dynamic worth watching as more chipmakers and robotics vendors compete for the same "who trains the robots" positioning.