On August 25, 2026, Skild AI officially released the S1 robot foundation model. According to the company's official blog, S1 can learn multi-step physical tasks lasting up to 10 minutes from a single human demonstration video—including pour-over coffee, repotting plants, and flipping pancakes, tasks that never appeared in the training data—and execute them in real time on real robots, without fine-tuning or any form of post-training.
One detail is most telling: when S1 first attempted to flip a pancake, the Skild team searched the entire pretraining dataset and found no pancake-flipping examples in it at all. S1 still inferred this out-of-distribution action from a single demonstration video and executed it successfully. This is not the result of engineering tuning; it is true generalization.
66% vs 9%: The Paradigm Gap Behind the Numbers
Skild's official blog released core performance data: using the same 100,000 hours of pretraining data, S1 achieved a 66% success rate on unseen tasks, while a comparable language-prompted vision-language-action (VLA) model achieved only 9%—a gap of more than 7x.
Another figure is even more direct: one video prompt is equivalent to approximately 380 task-specific post-training examples. For the robotics industry, this carries enormous practical significance—the cost of collecting robot training data is extremely high. Each hour of teleoperation data often requires prolonged coordination with specialized personnel, and Skild itself disclosed that its quality-control spending is 3x the data-collection cost. If a single video prompt can replace hundreds of specialized demonstrations, the cost of deploying new tasks will shift by an order of magnitude.
It should be noted that this 66% success rate comes from Skild's internal benchmark and has not yet been verified by independent third-party evaluators. In robot benchmarking, task configurations and environmental conditions have an outsized impact on results, and the generalizability of this figure still needs support from more real-world data.
Technical Route: Why Video Is Better Suited Than Language for Teaching Robots
Current mainstream robot foundation models, such as π0, RT-2, and OpenVLA, largely follow the prompting logic of language models: describe the task in natural language, then have the model generate action sequences. The natural limitation of this route is that language is inherently ambiguous. The sentence "put the cup on the table" does not contain the physical details of how to adjust the grip angle or at which moment to release the fingers.
S1 has chosen a different technical route. Skild's official blog describes it as "treating video as the program"—the model is centered on episodic data during pretraining, always specifying tasks through in-context demonstrations rather than language labels. This forces the model to learn, at a fundamental level, to infer the "demonstrator's intent," "functional correspondences," and "task progress," rather than memorizing statistical language-action mappings.
More critical is the behavioral difference after scaling. Skild's internal tests show that with a small amount of pretraining data (1,000 hours), language-prompted policies were even slightly better than in-context learning (53% vs 43%). But when the data scaled to 100,000 hours, the advantage of in-context learning grew explosively; language-prompted models degraded 3x faster than S1 under L5 conditions with significant distribution shift. This scaling advantage is highly analogous to the logic of emergent abilities in large language models as parameter counts grow.
Company Background: A $14.4 Billion Bet From a CMU Lab
Skild AI was founded in 2023 by Carnegie Mellon University (CMU) professors Deepak Pathak and Abhinav Gupta, both academic heavyweights in the field of robot learning. According to The Robot Report, the company completed a $300 million Series A round in July 2024, led by Lightspeed Venture Partners, Coatue, SoftBank, and Bezos Expeditions, at a valuation of $1.5 billion.
In January 2026, Skild completed a $1.4 billion Series C round led by SoftBank, with participation from Nvidia's NVentures, Samsung, LG, Schneider Electric, and Salesforce Ventures, at a valuation exceeding $14 billion. From its founding in 2023 to a $14 billion valuation: less than three years.
This funding trajectory itself reflects capital's assessment of the robot foundation model race: software-layer foundation models will become the focal point of competition for control of the next-generation robotics industry. Skild's strategic position is not to build robots, but to become "the platform that puts brains in other people's robots"—a logic similar to what Android meant for smartphones. It also means the competition it faces comes not only from full-robot manufacturers like Figure AI and Boston Dynamics, but also from giants like Nvidia and Google DeepMind that control training infrastructure.
The Real Anomaly Signal: Where Does the Generalization Come From?
The technical aspect of S1 most worth examining is not what it can do, but the source of its generalization capability.
The bottleneck of traditional robot learning has never been algorithms, but the fragility of data distributions—a model trained in the lab fails when the table changes or the lighting changes. S1's demonstrated "robustness to scene perturbations and object substitutions" (as stated in Skild's blog), and its occasional ability to "improve upon flawed demonstrations," point to a deeper capacity for environment modeling, rather than simple pattern matching.
Skild's pretraining data sources include teleoperation data, UMI robot data, first-person video, and simulation data. The scale of the simulation portion is described as "trillions of physical simulation episodes"—this is the key. Simulation data naturally possesses diversity and randomization; large-scale simulation pretraining is one of the underlying explanations for S1's ability to generalize from a single real-world video. This technical path follows the same lineage as DeepMind's simulation-accelerated training approach, but S1 has advanced it to the application layer of real-time in-context learning.
Uncertainties That Remain
The task types chosen for S1's launch demonstrations are worth noting: coffee, pancakes, plants—these are desktop manipulation scenarios emphasizing hand dexterity, not high-speed production lines, hazardous environments, or tasks requiring long-distance navigation planning. The currently publicized testing scope remains limited for assessing S1's performance in industrial-grade real-world scenarios.
Additionally, S1's 66% success rate is Skild's internal test data. In the field of robot benchmarking, task design, the definition of success criteria, and the choice of test environments can have a decisive impact on results; independent reproduction and third-party evaluation are necessary steps to verify the reliability of this figure.
Verdict
S1's significance is not that it can replace existing robot training pipelines, but that it has, for the first time, used concrete numbers to make one thing clear: on the foundation of sufficiently large-scale simulation pretraining, robot in-context learning can achieve a true leap beyond language-era VLA models.
The fundamental problem with language-prompted robot policies is that language itself cannot encode the fine-grained constraints of the physical world; the approach of using video demonstrations as "task programs" bypasses language as a lossy intermediate layer. S1's core technical bet is: scaled simulation data + in-context learning can achieve more efficient in-the-wild generalization than language-based VLAs. On this point, it has already provided convincing preliminary evidence on internal benchmarks.
For the robotics industry, the next question to answer is not "can in-context learning be applied to robots," but "under what conditions can the 66% success rate be achieved, and when can it advance to 90%+ production-grade reliability." This is the promise Skild must now deliver on in real customer scenarios, and the core test of whether its $14 billion valuation can hold.
Sources: - [Introducing S1: In-Context Learning for Robotics | Skild AI](https://www.skild.ai/blogs/s1) - [Skild AI Raises $1.4B to Build 'Omni-Bodied' Robot Brain - The Robot Report](https://www.therobotreport.com/skild-ai-raises-1-4b-building-omni-bodied-robot-skild-brain/) - [Skild AI S1 Model Learns 10 Minute Robot Tasks From One Video, Replacing 380 Examples | HuggingNews](https://huggingnews.com/ai/skild-ai-s1-model-learns-10-minute-robot-tasks-from-one-video-replacing-ce34c26a) - [Skild AI Presents S1 Model for One-Video Robot Tasks | Digg](https://digg.com/tech/rbnvscse) - [AI Startup That Builds a Brain for Robots Valued at $14 Billion | AI Business](https://aibusiness.com/robotics/skild-ai-startup-builds-robot-brain)© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接