AI news videos

One Video Teaches a Robot a New Task, Skild and NVIDIA Say

NVIDIA has published the detail of its role in Skild AI's S1 robot foundation model. Worth being precise about the timing: S1 itself launched the previous week. What NVIDIA published on September 10, 2026 is the infrastructure account, plus a factory deployment.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• HOOK. The narration says Skild says one video demonstration teaches a robot a new task with no retraining, and that NVIDIA just detailed how. The article supports the capability claim directly: it describes Skild AI's new S1 robot foundation model as "designed to learn previously unseen, long-horizon tasks from a single video demonstration" and says the model "uses video as input to understand and execute the task without updating its weights or undergoing task-specific post-training, a technique called in-context learning". The attribution wrapper "Skild says" is deliberate, because the capability description originates with Skild and reaches the reader through Skild's infrastructure vendor. The words "NVIDIA just detailed" are supported by the publication facts themselves: the piece is NVIDIA's own blog post, byline Sasa Docca, published September 10, 2026, and its stated purpose is to lay out how Skild used NVIDIA technology. SCENE 1. The narration says NVIDIA says Skild AI built its S1 model on NVIDIA infrastructure, spanning synthetic data, training, simulation and deployment, and that the model launched last week. This tracks the article sentence "Skild built S1 and conducted the research on NVIDIA AI infrastructure, part of a broader collaboration spanning synthetic data generation, model training, simulation and real-world physical AI deployment", and the subhead, which says the S1 robotic foundation model "harnesses NVIDIA technologies spanning synthetic data generation, model training, simulation and real-world deployment". The freshness statement is taken verbatim in substance from the article clause "The model, launched last week". Every framing in this script was chosen so that no viewer could conclude the model debuted today. What is new today is NVIDIA publishing the detail of its own infrastructure role, plus the Foxconn deployment. SCENE 2. The narration says an operator records a video of the task and uses it as the prompt, that no retraining follows, and that tasks can span dozens of steps. The article says "An operator records a video of the desired task and provides it to the model as a prompt. It interprets the demonstrated intent, objects and sequence, then maps them into actions for the robot in front of it, with no retraining, and often for a task not covered by its pretraining dataset." It also says "These tasks can span dozens of manipulation steps and require the robot to compose skills in sequences it hasn't previously performed." SCENE 3. The narration says Skild says the model succeeded on about 66 percent of steps, versus 9 percent for a similar system, and that Skild, NVIDIA and Foxconn deploy robots assembling Blackwell systems. The article says "In Skild's tests on new, multistep tasks, its S1 robot succeeded about 66% of the time at each step, compared with 9% for a similar AI system, a more than sevenfold improvement." The deployment sentence is "Skild, NVIDIA and Foxconn are deploying the Skild Brain on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems." The narration says "a similar system" and never names it, because the article never names it. SCENE 4. This is the honesty scene and it makes no new factual claim about the technology. It states whose numbers the previous scene reported, where they were published, that the comparison system is unnamed, and that there is no independent verification. ENDCARD. Standard Inboxsmith pitch, no factual claim about the story. HONESTY CAVEAT: this story rests on a single source and that source is not neutral. The article is NVIDIA's own corporate blog at blogs.nvidia.com, tagged "Customer Stories", "Physical AI" and "Industrial and Manufacturing". NVIDIA is the vendor whose hardware and software stack the piece promotes, and Skild AI is the customer and partner being profiled, so the piece functions as vendor marketing as well as news. Every performance figure in it is Skild's own number, produced by Skild's own tests, and published by its infrastructure vendor. That includes the 66 percent versus 9 percent per step success comparison, the claim of a more than sevenfold improvement, the estimate that one short video example is worth roughly 380 hands on training examples, the 50 to 100 hours a person would spend collecting those examples, the 11 minutes from demonstration to autonomous execution in a plant potting test, the 100 million dollar annual revenue run rate reached 10 months after first commercial deployment, and the more than 60 deployment partnerships. For none of these does the article give a methodology, a sample size, a task list, a date range, or an independent replication. The 9 percent baseline is attributed only to "a similar AI system", which is not named, so that comparison cannot be checked by anyone outside the two companies. The success rate is also reported per step rather than per completed task, which is a much weaker measure than end to end task success on a job with dozens of steps, and the article does not report end to end success at all. On timing, the model itself launched last week, not today. Today's news value is NVIDIA detailing its infrastructure role and the Foxconn factory deployment. Research support was thin: YouTube research returned zero of five usable transcripts because the local yt dlp build is out of date, and the timeline probes returned only unrelated videos including Fortnite clips, memes and a TED talk. No third party reporting and no independent verification of any kind was available or used, so this article is the sole source for the entire script. Facts deliberately left out of narration, for the 132 word limit and to avoid stacking unverified vendor supplied figures into a 60 second video: the 100 million dollar annual revenue run rate and the more than 60 deployment partnerships; the roughly 380 training examples equivalence and the 50 to 100 hours of manual collection; the 11 minute plant potting result; the more than sevenfold framing of the 66 versus 9 percent gap; the list of demonstrated tasks such as plant potting, pancake making, pour over coffee brewing and kit assembly; the claim that unfamiliar tasks can last up to 10 minutes; the specific Foxconn workflow of installing a busbar and limit block and fastening 16 screws; the quote from Skild cofounder and CEO Deepak Pathak; and the full list of named NVIDIA products, which includes Cosmos, Cosmos Curator, Omniverse, Isaac Sim, Isaac Lab, the Newton physics engine, Nsight and TensorRT. Several phrasings in the article are figurative and were rewritten literally rather than repeated, including "from the lab into factories", "step change", "breaks the cycle", "already in action on the factory floor" and "shared robot brain".(NVIDIA's own first party announcement, "Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video", published on blogs.nvidia.com on September 10, 2026)