Skild AI Unveils S1 Foundation Model, Enabling Robots to Learn 10-Minute Tasks From a Single Video

When a large language model needs to learn a new task, users do not retrain its neural network; they simply type a new instruction into the prompt box. Skild AI’s newly unveiled S1 foundation model applies that exact mechanism to physical robotics, with one crucial difference: the prompt is a single video of a human doing the job. By treating a video demonstration as an immediate instruction rather than the starting point for a lengthy data-collection cycle, the system attempts to bypass the primary bottleneck in robotic automation.[3]

Announced in late August 2026 and highlighted today by NVIDIA as a key deployment of its physical AI infrastructure, S1 is designed to execute long-horizon manipulation tasks it has never seen before. The model requires no task-specific fine-tuning and no weight updates. Instead, it relies entirely on in-context learning, a paradigm shift that moves robotics closer to the flexibility of modern text generators.[1][2]

Historically, teaching a robot a new skill has been a grueling exercise in data collection. To get a robotic arm to fold a shirt or sort a new type of package, engineers typically collect tens to hundreds of hours of teleoperation data—where a human manually drives the robot through the motions—and then post-train a specialist policy. This task-by-task retraining is expensive, slow, and limits robots to highly structured, repetitive environments.[2][3]

Skild AI argues that this data bottleneck is what keeps robotics stuck in the lab. S1 bypasses the post-training phase entirely. The robot is shown an egocentric video of a human performing the task. That video is loaded into the model's context window, and the frozen neural network weights map the human's intent, functional correspondences, and spatial movements directly to the robot's motor controls.[2]

“Learning by experience, and not preprogramming, is the step change that has happened in robotics,” said Deepak Pathak, cofounder and CEO of Skild AI. The company notes that in meta-learning terms, the model's massive pre-training acts as an outer loop that teaches the policy how to learn from context. At inference time, the video demonstration drives the inner loop, dictating the robot's actions without changing a single parameter in the model.[1][2]

The tasks S1 can handle are not brief, three-second pick-and-place motions. Skild demonstrated the model executing jobs lasting up to 10 minutes across dozens of manipulation steps. Examples included repotting a plant, making pour-over coffee, assembling a kit, and cooking pancakes. These are sequences that require the robot to compose atomic skills in ways it has never previously performed.[1][2]

In the plant-potting demonstration, only 11 minutes passed between the human recording the video demonstration and the robot beginning its autonomous execution on the hardware. During that execution, the robot demonstrated novel test-time behaviors that were absent from its pre-training, such as digging into the soil to make room for the plant and pressing a paper filter into a coffee funnel.[1][2]

S1 also showed signs of common-sense physical understanding and error recovery. In one internal test, the human demonstration showed a person watering a plant with a watering can, but the robot was only provided with a cup of water. The model adapted to the discrepancy and successfully used the cup. In another example, the prompt showed a glass being filled with juice, but the physical glass in front of the robot was already nearly full; S1 adjusted and simply topped it off.[2]

To quantify the model's performance, Skild ran controlled internal benchmarks comparing S1's video-prompted approach against a standard language-prompted Vision-Language-Action model. At 100,000 hours of pre-training data, the in-context model achieved a 66% average step-success rate on unseen tasks, compared to just 9% for the language-conditioned baseline—a more than sevenfold improvement.[1][2]

The company estimates that matching S1’s single-video performance through traditional methods would require roughly 380 teleoperated demonstrations, equivalent to dozens of hours of manual data collection. This suggests that a single video prompt can effectively replace weeks of engineering work, fundamentally altering the economics of deploying robots in dynamic environments like manufacturing floors and warehouses.[1]

The model's development relies heavily on NVIDIA's accelerated computing stack. Skild uses NVIDIA Omniverse and Isaac Lab for physics simulation and synthetic data generation, allowing the model to experience diverse scenarios before touching physical hardware. “NVIDIA Isaac Lab and NVIDIA Cosmos technologies help Skild create the scalable, diverse experience its robots need to learn across many scenarios and embodiments,” Pathak noted.[1]

This infrastructure has supported rapid commercial traction. According to NVIDIA, Skild AI reached a $100 million annual revenue run rate just 10 months after its first commercial deployment. The company has built more than 60 deployment partnerships spanning manufacturing, logistics, inspection, security, and food preparation, indicating that the demand for adaptable, omni-bodied robot brains is substantial.[1]

However, the 66% success rate comes with caveats. It is a step-success metric that includes human recovery interventions, meaning it does not guarantee the robot can complete an entire 10-minute job unassisted 66% of the time. Furthermore, S1 is not yet an open API or a downloadable set of weights, meaning independent researchers cannot yet verify the benchmarks. The true test will be whether the model's in-context learning holds up outside controlled laboratory environments when faced with the chaotic lighting, friction, and wear of a real factory floor.[2][3]