Skild AI is a robotics company, founded in 2023, that builds one AI model, the Skild Brain, meant to control any kind of robot: arms, quadrupeds, humanoids and mobile manipulators. It doesn’t sell its own consumer robot. It sells the software to businesses, often running on robots from partners such as ABB and Universal Robots. As of October 19, 2026, it’s valued at over $14 billion and says it earns more than $100 million a year.
There is no public download, API sign-up or price list on Skild’s website. If you want to use the Skild Brain, you go through Skild or one of its robot-maker partners.
What the Skild Brain is
The Skild Brain is a robot foundation model: a single neural network trained on many robots and tasks that can be adapted to new ones. Skild’s word for this is “omni-bodied.” In a September 2025 post, the company said it trained the model across a simulated set of 100,000 different robot bodies so it wouldn’t depend on any one robot’s shape.
According to Skild’s Series C announcement, the model learns from four kinds of data:
- Simulation: large numbers of synthetic trials in approximate physics simulators.
- Internet video: human videos of people handling objects, used to teach how things are manipulated.
- Teleoperation: people controlling robots remotely, which records images, joint positions and torques together.
- Real deployments: data from robots already working for customers.
The first two build general skill; the last two adapt it to real tasks. The pitch is that every new deployment adds data that improves the model for all customers. That claim comes from Skild and hasn’t been independently measured.
What S1 is
S1 is Skild’s flagship model, announced on August 18, 2026. Its selling point is in-context learning: you show it a video of the new task as a prompt, and it attempts the task without retraining. Skild compares this to the shift from BERT to GPT-3 in language models, where users stopped fine-tuning for every task and started prompting instead.
That matters because the normal routine in robot learning is slow. A new task usually needs hours of teleoperated demonstrations and a fine-tuning run before the robot does it reliably. Skild’s S1 results cover tasks lasting up to about 10 minutes from a single video prompt. These are Skild’s own tests; no outside group has published a reproduction.
Skild followed it with a physical self-play post on September 23, 2026, describing a post-training method in which S1 learns fast, dynamic skills such as soccer by competing against copies of itself in simulation. Skild presents it as a way past the limit of learning only from human data.
Who’s paying for it
In a September 10, 2026 post, co-founders Deepak Pathak and Abhinav Gupta wrote that Skild had passed $100 million in annual recurring revenue ten months after its first commercial deployment. They gave these details:
- Customers: more than 60 paying customers, in moving goods, deliveries, site inspection, security, food preparation, warehouses, factories and data centers.
- Revenue mix: about 90% from manipulation tasks and 10% from mobility. Warehouse robots from the Zebra (Fetch Robotics) acquisition make up 4%.
- Named projects: dual-arm robots assembling NVIDIA Blackwell systems with NVIDIA and Foxconn; work toward automating wire-harness manufacturing at Sumitomo Wiring Systems; and a pilot of S1-powered robots in commercial kitchens with Mitsui & Co.
The figures are self-reported. Skild is private and doesn’t file public financial statements, so there is no outside audit to check them against.
Partners and acquisitions
Skild’s route to market runs through existing robot makers rather than its own hardware:
- March 19, 2026: partnerships with ABB Robotics, Universal Robots and MiR, with NVIDIA, to deploy the Skild Brain on their robots. The idea is to fine-tune the base model on a small amount of task data instead of programming each task by hand. Our Korean analysis of where a general robot brain meets industrial safety controls looks at what that split means on a factory floor.
- April 15, 2026: acquisition of Zebra Technologies’ robotics business, formerly Fetch Robotics, including its warehouse robots and orchestration software. We covered why Skild bought it separately.
Funding and valuation
Skild has raised money in two big rounds that it announced itself:
- July 2024: a $300 million Series A led by Lightspeed, Coatue, SoftBank and Jeff Bezos.
- January 14, 2026: a $1.4 billion Series C led by SoftBank, with NVIDIA’s NVentures, Macquarie Capital and Bezos Expeditions, valuing the company at over $14 billion. Strategic investors include LG, Schneider Electric, CommonSpirit and Salesforce Ventures.
Skild hasn’t announced a later round as of October 19, 2026. Its shares aren’t publicly traded.
How it compares with similar companies
Skild sits in the same group as Physical Intelligence and Google DeepMind’s Gemini Robotics: companies building a general model rather than one robot. The practical differences are in access. Physical Intelligence publishes open models through openpi and has started an API beta; Google offers one Gemini Robotics model through the Gemini API. Skild offers neither and sells only through business deployments. Our guides to Physical Intelligence and openpi and Gemini Robotics show what you can try yourself.
For where Skild fits among robot makers, chip suppliers and simulation companies, see our map of physical AI companies.