Skild AI says one video is now enough to teach a robot a 10-minute task -- in a benchmark the company designed and ran itself

Skild AI introduced a new robot foundation model this week that it says needs only a single video of a person doing a task -- not the thousands of hours of task-specific training data most robot-learning systems require -- to reproduce that task on a physical robot with no fine-tuning. The company calls it S1, and its headline number is a 66% success rate on tasks the model was never trained on, drawn from ten-minute-long, multi-step jobs like potting a plant, frying a pancake, making pour-over coffee and assembling a kit. That number, and every other number in this piece describing what S1 can do, comes from Skild's own blog post. No outside lab has run the test.

The pitch is a genuine shift in approach, whatever the eventual number turns out to be under independent testing. Most robot-learning systems today are language-conditioned: told what to do in words, then fine-tuned on many recorded demonstrations of that specific task before they can reliably perform it. S1 is what Skild calls an in-context learner -- shown one example, it is expected to generalize the way a large language model generalizes from a single prompt, without any additional training step. Skild's own comparison: one in-context video example produced results roughly equivalent to 380 traditional training episodes for tasks the model had never seen -- a claim, if it holds up outside Skild's own testing, that would meaningfully cut the data-collection cost that has made general-purpose robot software slow and expensive to build.

Skild's own numbers show a large gap between the model's best and worst reported conditions. At the top end of its training-data scale -- 100,000 hours of pre-training -- S1 reports a 66% success rate on tasks it has never seen, against a language-conditioned baseline system Skild built for comparison, which reached only 9% success under the same conditions. On tasks the model has seen before, at a much smaller pre-training scale of 1,000 hours, in-context learning reached 43% success and outperformed the language-conditioned approach as the data scale grew. Skild frames the gap as evidence that in-context learning improves faster with scale than language conditioning does; an outside reader has no way yet to check whether that pattern holds outside the specific tasks and conditions Skild selected to test it on.

The company also disclosed an operational detail that says as much about its ambitions as the success-rate numbers: Skild reports spending roughly $3 on quality screening for every $1 spent on data collection, a ratio that suggests the model's performance is more sensitive to demonstration quality than to raw volume of video. That tracks with the in-context learning pitch generally -- if one good example is meant to substitute for hundreds of training episodes, a bad example is expensive in a way a large, noisy dataset usually isn't.

Skild is not a new company making its first claim. Founded in Pittsburgh in 2023 by Deepak Pathak and Abhinav Gupta, it has spent two years pitching what it calls an "omni-bodied" robot brain -- one model retrofittable across different physical robots and tasks, including, per the company's own description, the ability to keep controlling a robot that has lost a limb or jammed a wheel without retraining. Those are the kind of claims that are hard to check from a blog post and easy to repeat in coverage that takes the company's framing at face value, which is largely how S1's launch has been covered so far: strong on the demo, thin on any outside party actually running the numbers.

Nvidia is both an investor in Skild's January round, through its NVentures arm, and a technology partner -- Nvidia's own published case study on Skild describes the startup using Nvidia's Isaac Lab simulator and Cosmos Transfer tools to generate training scenarios across thousands of simulated robot instances, part of what Skild's CEO credits with getting hardware costs for a capable robot down to $4,000-$15,000, against $250,000-plus for traditional systems. That puts Nvidia on both sides of any claim about how well Skild's software runs: an investor with a stake in the valuation, and the vendor whose own marketing benefits from Skild's success story. Skild's other named competitors in general-purpose robot software include Qualcomm, Sanctuary AI and X Square Robot, none of which have published a directly comparable single-video, in-context benchmark; the closest thing to a rival data point right now is the language-conditioned baseline Skild built and tested itself.

The valuation trajectory is real and independently reported, even if the underlying capability claims aren't: Skild raised $300 million in 2024, then $1.4 billion in a SoftBank-led Series C closed in January 2026 that valued the company above $14 billion -- more than triple the $4.5 billion figure from a round seven months earlier, according to Bloomberg and TechCrunch. CEO Deepak Pathak told Bloomberg the company has now raised more than $2 billion total. That funding history is well-sourced and confirmed by multiple independent outlets. S1's performance numbers are not the same kind of fact -- they are seven months newer than the last figure anyone outside the company has verified, and they come from the company that built the product being measured.

None of that makes S1 a bad model or Skild a dishonest company -- it makes it an unverified one, on this specific release, which is a different problem with a different fix: someone other than Skild needs to run the test. Here is the strongest version of that case.

If S1's numbers hold up outside Skild's own testing, the practical effect would be real: robot deployment currently bottlenecks on collecting and labeling task-specific training data, and a model that generalizes from one video instead of hundreds of hours would cut that cost sharply for any company trying to put a robot into a new task quickly. That is exactly why a company under pressure to justify a valuation that has tripled in seven months has every incentive to publish the most favorable version of this result -- which is not an accusation of dishonesty, just a reason to wait for someone other than Skild to run the test before treating 66% as the real number rather than the best one Skild found.