Skild AI Introduces S1: Robot Expected to Learn a New Task from a Single Video
Skild AI’s S1 model is intended to enable robots to perform previously unknown multi-step tasks after a single video demonstration without additional training. However, the publicly available results currently come from the company.
Skild AI S1 is a new robotics foundation model designed for in-context learning, which the company introduced on August 18, 2026. According to Skild AI, a robot only needs to be shown a task in a video: the model is then expected to perform it without changing its own weights and without additional training. On September 10, NVIDIA described how Skild uses its infrastructure and software for physical AI in development and its stated deployment in manufacturing.
The basic idea is to replace some conventional programming or the collection of extensive training data with a demonstration. The video serves as a task example. Skild says the model does not change its parameters during execution, meaning it does not undergo new post-training for each individual task.
Skild AI S1 Relies on a Video Demonstration of the Task
Skild AI demonstrated S1 on tasks that, according to the company, the model had not previously seen and that may take up to ten minutes. The examples cited include repotting a plant, preparing pancakes and coffee, and assembling a kit. These are multi-step activities in which the robot needs to connect individual actions in the correct order and work with objects in its environment.
The company also published its own comparisons. It claims a 66 percent success rate on new multi-step tasks, a sevenfold improvement over a comparison system, and that one demonstration may correspond to approximately 380 training episodes. However, these figures should be viewed as Skild AI’s internal results. No independent evaluation or public technical paper is available to allow the methodology and results to be reproduced.
Skild and NVIDIA materials likewise discuss the model’s resilience to disturbances, ability to correct errors, and readiness for broader manufacturing use. These properties are not yet supported by independent benchmarks in publicly available materials.
NVIDIA Describes the Technology Behind It and a Manufacturing Use Case
In its September article, NVIDIA confirmed that Skild uses its AI infrastructure, Cosmos models, the Omniverse platform, and the Isaac Sim simulator.
The article also describes work by Skild, NVIDIA, and Foxconn on high-precision assembly of Blackwell systems. However, this was not a new partnership announced on September 10. Skild AI had already publicly announced its collaboration with NVIDIA and Foxconn on Blackwell assembly on March 19, 2026.
NVIDIA’s September communication therefore primarily provides a closer description of the technology infrastructure being used and the manufacturing scenario surrounding S1. The available information does not indicate the specific scale or timing of deployment on Foxconn’s production lines.
What Still Needs to Be Verified
Learning from a single demonstration is particularly interesting for robotics in tasks that change frequently and where creating a separate program or training set is difficult. In the case of S1, however, the public materials so far rely mainly on presentations by the manufacturer and its partner.
Important signals will include independent benchmarks, a technical publication, or information about the model’s availability to customers. In manufacturing scenarios, specific confirmation of stable success rates, cycle speed, and safety during long-term operational deployment will also be important. Details about the scale and schedule of use in Foxconn factories are likewise missing.
Sources
- Skild AI — Introducing S1: In-Context Learning for Robotics – Confirms the introduction of S1, the video-prompting principle without changing weights, and internal results including tasks lasting up to 10 minutes.
- NVIDIA Blog — Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video – Confirms the use of NVIDIA technologies and describes the stated manufacturing use with Foxconn.
- Skild AI — The Reindustrial Revolution: Partnering with ABB Robotics, Universal Robots, and NVIDIA – Documents that the collaboration with NVIDIA and Foxconn on Blackwell assembly was announced as early as March 19, 2026.
Verified and updated: 09/11/2026 15:21
