Skild AI unveils S1 robotics model that learns physical tasks from a single video

Teaching a robot to fold laundry used to require painstaking programming, thousands of demonstrations, and a small army of engineers. Skild AI thinks a single video should do the trick.

The Pittsburgh-based startup just launched S1, a robot model that can learn physical tasks from watching one video of a human performing them. No fine-tuning, no hardware-specific adjustments. Just watch and do.

How S1 actually works

Skild’s underlying technology, called Skild Brain, uses a hierarchical architecture split into two layers. The high-level policy handles the big-picture stuff: understanding what task needs to happen and planning the general approach. The low-level controller translates that intent into actual motor commands, the precise joint angles and force vectors that make a gripper close around a cup without crushing it.

Skild Brain was trained on trillions of simulated physics episodes and millions of human action videos. When S1 watches a new video, it’s not starting from scratch. It’s mapping what it sees onto a deep reservoir of physical intuition it already possesses. The company reports that in real-world tests, robots running S1 achieved task completion rates between 60% and 80% within hours of initial data collection.

Skild claims its model needs less than one hour of targeted robot data to pick up a new skill from video observation.

A startup growing at warp speed

Skild AI was founded in May 2023 by Deepak Pathak and Abhinav Gupta, both Carnegie Mellon University researchers who saw an opportunity to build a general-purpose brain for robots rather than the task-specific systems that have dominated the field for decades.

In mid-2024, Skild raised a $300 million Series A at a $1.5 billion valuation. Then in January 2026, SoftBank led a roughly $1.4 billion investment that pushed Skild’s valuation north of $14 billion. That’s nearly a 10x increase in valuation in about 18 months. The investor roster also includes Amazon and NVIDIA.

Why general-purpose matters

Skild’s pitch is that a single foundational model can control humanoids, manipulators, mobile platforms, and other robot form factors without needing to be retrained from the ground up each time.

The 60% to 80% task completion rate reveals the gap that still exists. In a manufacturing context, 80% accuracy means one in five attempts fails. That’s a problem if the task involves expensive components or safety-critical operations.

With $1.7 billion in total funding and a valuation that’s climbed from $1.5 billion to over $14 billion in roughly a year and a half, Skild has the resources to make a serious run at this problem.