Skild AI S1 Model Learns Robot Tasks From a Single Video Demo

A robot model you teach with one video

Skild AI, a robotics startup based in Pittsburgh, has released S1, which it calls its flagship robot foundation model. The company's main claim is that S1 can carry out a manipulation task it has never seen after watching one video of that task, with no fine-tuning and no task-specific post-training.1113 The video works the way a prompt works for a chatbot. It goes into the model's context window, and the same frozen weights then try to turn the demonstrator's intent into actions that suit the robot's own body and its surroundings.1117 Skild says those same weights produced every example it published.15

The showcase tasks were potting a plant, cooking pancakes, making pour-over coffee and assembling a kit. Each runs for up to ten minutes and involves dozens of manipulation steps.1112 The plant-potting demo got the most attention. Skild says 11 minutes passed between the start of the human recording and the robot beginning the task on its own.11 One detailed account gives the timeline: recording ran from 9:16 p.m. to 9:22 p.m., and the robot started at 9:27 p.m.20 Skild also says S1 is already in use with its commercial partners and has opened sign-ups for early access.1115

Reports disagree on the launch date. One outlet puts it at August 18, 2026.15 Most others, including Skild's own post on X, tie the release to August 25.1416 Trade coverage was still writing up the launch in mid-September.11

The numbers, and what they measure

The most-quoted figure comes from a controlled comparison Skild ran. It trained two models on the same data, architecture and compute. One was told what to do with a video, the other with text.19 After pre-training on 100,000 hours of data, the video-prompted S1 reached 66% success on unseen tasks. The language-prompted vision-language-action model reached 9%.11 On tasks the models had already seen, the gap was small: about 96% versus 89%.25 Skild's view is that language holds up for familiar work but cannot describe a genuinely new physical skill. Telling a robot to "flip the pancake" leaves out the motion that a video shows directly.25

The second figure concerns efficiency. By Skild's estimate, one video demonstration matched a conventional model fine-tuned on about 380 demonstrations. Collecting that much data for long tasks took 50 to 100 hours of teleoperation.11 Many people summed this up as one video beating 380 demos, but that leaves out two caveats.

First, the 66% figure counts the share of steps completed, averaged across a task. It is not the share of fully autonomous runs. One analysis notes that humans reset the robot after some failures.18 Second, the 380-demonstration figure was estimated by interpolating between experimental data points. It should not be read as a general exchange rate.18 One report also notes that the conventional fine-tuned policy eventually did better than S1. With 2,000 demonstrations it reached 86%, against S1's 66% from a single prompt.20 So the case for S1 is speed and cost, not the highest possible accuracy. It gets to useful performance on a new task in minutes, where the conventional route takes days of data collection.

Skild also tested how far a scene can change before S1 fails. The model reportedly coped when objects were moved by 30 cm and rotated 45 degrees, or replaced with different objects that serve the same purpose. Performance dropped sharply only when the new layout called for a different plan, such as using the other arm.25 Some behaviors appeared without targeted training, according to the company. The robot retried after failed attempts, used a cup when no watering can was available, and corrected mistakes made by the human demonstrator.2019 After the robot flipped a pancake, Skild searched its pre-training data for pancake flipping and says it found none.20

Where the coverage agrees, and where it differs

Outlets agree on the facts but read them differently. Enthusiastic accounts describe S1 as robotics' version of the prompting shift in large language models. One commentator argued that if the approach scales, the large cost of data is paid once during pre-training and then spread across thousands of new tasks.14 Skild encourages this comparison. It has described robotics as stuck in a "BERT era" in which every new skill needs its own training run.25

More cautious coverage points out that all the reported numbers come from Skild's internal benchmarks. No independent benchmark has been published, and the model weights are not open for outside researchers to test.1816 One builder-focused guide adds that it is unclear which robot types S1 supports. The featured tasks used a two-armed manipulation setup, not a humanoid.1625

There is a smaller mix-up as well. One explainer says S1 achieves 60–80% task completion in early real-world tests.16 That range, however, appears in an earlier Nvidia case study about the broader Skild Brain, where it describes performance within hours of data collection.23 Readers should not treat it as an independent check of S1.

The cautious reading is the correct one. The idea behind S1 is more significant than any single number. The model treats a demonstration as a statement of the goal, not a path to copy exactly. It learns how to learn from demonstrations during pre-training, then adapts at run time without changing its weights.1920 That is a real change in how robot learning is set up. Whether it is reliable enough for factories and warehouses is still unproven.

How S1 grows out of the Skild Brain

S1 continues a bet Skild has been making publicly for more than a year. The company is building what it calls an "omni-bodied" brain, a single model meant to control humanoids, quadrupeds, tabletop arms and mobile manipulators.2428 Its central argument is that robotics has no equivalent of the internet's text data. Instead of relying mostly on teleoperation, Skild pre-trains on physics simulations and on online videos of humans doing tasks.2421 One explainer describes the pre-training mix as trillions of simulated episodes plus millions of human action videos.16 Another source says S1's pre-training also includes teleoperation data and data captured with gloves.17

In January, Skild showed the Skild Brain learning household skills from human video with less than an hour of robot-specific data. It also pointed to in-context learning as the source of its adaptability.21 Earlier work focused on recovering from hardware damage. The company said the model recovered from jammed wheels within two to three seconds and handled payloads up to 1.5 times the robot's body weight.23 S1 applies the same in-context approach to long manipulation tasks.1620

Nvidia's infrastructure is closely involved. S1 was trained on Nvidia AI infrastructure.11 In September, Nvidia described the stack Skild uses, which includes Cosmos, Omniverse, Isaac Sim and Isaac Lab.18

Skild is not the only company working on video-taught robots. 1X has built a world model trained on first-person human video. Mentee has shown its humanoid learning to swap a battery from one human demonstration.27 Skild argues that most rival in-context work covers either short tasks or tasks similar to the training data.16 It also says that vision-language models with a small amount of robot data added are a "Potemkin village."25 That criticism will only carry weight if S1's results are reproduced outside Skild.

The money behind the model

S1 comes from one of the best-funded companies in embodied AI. In January 2026, Skild raised about $1.4 billion in a Series C led by SoftBank, at a valuation above $14 billion. Nvidia's venture arm, Macquarie, Bezos Expeditions and 1789 Capital also invested, along with strategic backers including Samsung, LG, Schneider Electric, CommonSpirit and Salesforce Ventures.3 Reports said the valuation had more than tripled from about $4.5 billion only seven months earlier.1 Chief executive Deepak Pathak said the company had raised more than $2 billion in total.1

Funding trackers disagree with each other. Estimates of total funding range from about $1.7 billion to $2.2 billion.267 One profile gives a $15 billion valuation from a round dated to March.2 Accounts of the Series B also differ, at either $135 million or $500 million.57 Disclosed revenue is much smaller: about $30 million in 2025, which implied a very high revenue multiple at the Series B.9

That gap between valuation and revenue explains why S1 matters commercially. One comparison puts Skild above software-focused rival Physical Intelligence on valuation and below hardware-first Figure AI.5 Skild's pitch to investors is that the general robot brain, more than any particular robot body, is where lasting value will sit.4 If customers can teach new tasks in minutes from one demonstration, robot deployments become cheaper and new tasks no longer require a new data-collection project. That is the commercial argument for a $14 billion software company.

What to watch

The next important step is evidence, not another demo video. Analysts are asking for published protocols, independent results, full-task success without human help, tests across different robots, and safety data for robots working near people.18 Skild has said it will publish more about how S1 was trained.12 Until outside results exist, S1 is the strongest evidence so far that robots can be prompted like language models, and it rests entirely on Skild's own measurements.