NVIDIA says Skild AI’s S1 learns robot tasks from one video

NVIDIA said on September 10 that Skild AI’s S1 robot foundation model is designed to learn an unfamiliar, long-horizon task from a single video demonstration. The claim rests on company evaluations, not independently verified production deployments.

If the approach generalizes, operators could adapt robots to new work with less task-specific data collection and fine-tuning. Available sources do not quantify savings in programming labor, integration effort, commissioning time or deployment cost.

In Skild AI’s description of S1, a video demonstration serves as an in-context task prompt that the model uses to generate robot actions. Skild AI says the process requires no task-specific fine-tuning or post-training, and that every launch example used the same model weights.

The company showed S1 attempting four tasks it described as previously unseen: potting a plant, cooking a pancake, making pour-over coffee and assembling a kit. It said the demonstrations included runs of up to 10 minutes and dozens of manipulation steps.

In an internal controlled comparison using 100,000 hours of pretraining data, Skild AI reported a 66% success rate for demonstration-prompted S1. A language-prompted policy using matched data and compute, and the same architecture apart from prompt embedding, scored 9%, according to the company. The published material does not provide independently validated full-task completion, intervention or failure rates.

Skild AI estimated that one in-context demonstration was equivalent to roughly 380 task-specific post-training demonstrations on its metric. The company said collecting that many long-horizon demonstrations required an estimated 50 to 100 hours of teleoperation. The comparison was interpolated, and task-specific post-training reached a higher company-reported score of 86% with 2,000 demonstrations.

Skild AI says S1 was built on NVIDIA AI infrastructure. In a separate robotics announcement, NVIDIA said Skild AI uses Cosmos world models for data generation and Isaac simulation tools to validate policies in simulation. For context on NVIDIA’s broader operational-AI strategy, DataPhoenix has covered NVIDIA’s separate operational-AI partnership with Palantir.

NVIDIA says Skild has built more than 60 deployment partnerships and names Foxconn work on Blackwell production lines; a separate NVIDIA announcement identifies ABB Robotics and Universal Robots as partners. Those published materials do not provide production-scale reliability, safety, intervention or failure-rate data for S1.