Skild AI Launches S1 Robot Model Trained on NVIDIA Physical AI Stack
Skild AI says its new S1 robot foundation model learns unfamiliar, long tasks from a single video, built using NVIDIA's simulation and training technologies.
ひとことで言うと
What did Skild AI announce with NVIDIA about its new S1 robot model?
Skild AI launched S1, a robot foundation model that learns new, multistep tasks from a single video demonstration without retraining. Built on NVIDIA's Isaac Lab, Cosmos, Omniverse and TensorRT technologies, Skild says S1 succeeded at 66% of steps on new tasks versus 9% for a comparable system, and the company has reached a $100 million annual revenue run rate.
要点
- Skild AI says its S1 robot foundation model can learn a previously unseen, long-horizon task from a single video demonstration, without updating its weights or undergoing task-specific retraining.
- The company reports S1 succeeded at about 66% of steps on new multistep tasks in its tests, compared with 9% for a similar AI system it compared against.
- Skild built and trained S1 using NVIDIA technologies including Isaac Lab, Cosmos, Cosmos Curator, Omniverse, Isaac Sim, Nsight and TensorRT, spanning data generation, simulation, training and deployment.
- Skild says it reached a $100 million annual revenue run rate roughly 10 months after its first commercial deployment and has more than 60 deployment partnerships.
- Skild, NVIDIA and Foxconn are deploying the Skild Brain on dual-arm manipulators to assemble NVIDIA Blackwell systems, according to the announcement.
Skild AI has released S1, a robot foundation model the company says can learn an unfamiliar, multistep task from a single video demonstration without retraining or additional task-specific data. The model, which launched last week according to NVIDIA, was built and trained using NVIDIA's physical AI infrastructure across data generation, simulation, training and deployment.
How S1 Works
Skild says an operator records a video of a task and feeds it to S1 as a prompt. The model interprets the objects, intent and sequence shown in the video, then maps that information onto actions for whichever robot it is controlling, without updating its weights. The company calls this approach in-context learning.
According to Skild, S1 can carry out unfamiliar tasks lasting up to 10 minutes, including plant potting, pancake making, pour-over coffee brewing and kit assembly, some involving dozens of manipulation steps the robot hasn't previously performed in that order. In one plant-potting test cited by the company, the team went from recording a demonstration to autonomous execution on hardware in 11 minutes. Skild also says the model can adjust when objects move and recover from errors during a task.
Skild reports that in its tests on new, multistep tasks, S1 succeeded at about 66% of steps, compared with 9% for a similar AI system, and estimates that one short video example can provide value comparable to roughly 380 hands-on training examples, which the company says would otherwise take a person 50 to 100 hours to collect manually.
NVIDIA Technology Behind the Model
According to NVIDIA, Skild trained its "shared robot brain" using simulation, human video and teleoperation data, drawing on several parts of NVIDIA's stack. Cosmos, described as an open world foundation model, is used to diversify training data and convert video into structured descriptions, while Cosmos Curator annotates, filters and organizes data at scale.
Skild also uses NVIDIA's Omniverse libraries and the Isaac Sim framework to generate data and validate robot behavior in simulated environments before real-world deployment, and applies reinforcement learning in Isaac Lab, an open modular robot learning framework powered by the Newton physics engine, to model forces, contact, collision and pressure. NVIDIA said it and Skild are jointly developing new GPU-accelerated simulation solvers for modeling how robots grip and manipulate objects, which will be added to Newton for all developers. NVIDIA Nsight tools are used to identify training performance bottlenecks, and the TensorRT software development kit optimizes the model for inference.
Commercial Deployment
NVIDIA states that Skild reached a $100 million annual revenue run rate about 10 months after its first commercial deployment and has established more than 60 deployment partnerships spanning manufacturing, logistics, inspection, security and food preparation.
One active deployment, described in the announcement, involves Skild, NVIDIA and Foxconn using the Skild Brain on dual-arm manipulators for assembly of NVIDIA Blackwell systems. In a demonstrated workflow, a robot installs a busbar and limit block and fastens 16 screws while adapting to disturbances during the task, according to NVIDIA.
Deepak Pathak, cofounder and CEO of Skild AI, said in the announcement: "Learning by experience, and not preprogramming, is the step change that has happened in robotics. NVIDIA Isaac Lab and NVIDIA Cosmos technologies help Skild create the scalable, diverse experience its robots need to learn across many scenarios and embodiments."
Source: NVIDIA Blog, "Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video," published September 10, 2026.
よくある質問
- What is Skild AI's S1 model?
- S1 is a robot foundation model from Skild AI that the company says can learn and perform a new, multistep task after being shown just one video demonstration, without retraining.
- How does S1 learn new tasks?
- According to Skild, an operator records a video of a desired task and gives it to the model as a prompt; S1 interprets the objects and sequence shown and maps them to actions for the robot in front of it, a technique the company calls in-context learning.
- What NVIDIA technologies did Skild use to build S1?
- Skild says it used NVIDIA Cosmos and Cosmos Curator for data, Omniverse and Isaac Sim for simulation, Isaac Lab with the Newton physics engine for reinforcement learning, and Nsight and TensorRT for training and inference optimization.
- Where is the technology being deployed?
- Skild states that it, NVIDIA and Foxconn are using the Skild Brain on dual-arm manipulators to assemble NVIDIA Blackwell systems, in addition to more than 60 other deployment partnerships across manufacturing, logistics, inspection, security and food preparation.