A robot that watches one demonstration, then tackles a task it has never seen before, is pushing robotics into a new phase. Skild AI built its S1 robot foundation model on NVIDIA AI infrastructure, giving the system a way to learn from experience instead of relying on task-by-task programming.
Skild introduced S1 in a research post on August 18, 2026, and NVIDIA detailed the collaboration on September 10, 2026. The announcement arrives alongside a major business milestone: Skild AI reached a $100 million annual revenue run rate 10 months after its first commercial deployment.
S1 Turns Video Demonstrations Into Robot Prompts
S1 is built as an in-context learner. An operator records a video of a desired task and provides that video to the model as a prompt. S1 then executes the task without updating its weights or undergoing task-specific post-training.
That design changes what the robot needs before it can act. Instead of receiving a new training cycle for every task, S1 uses the demonstration as a goal specification. The system does not simply copy a fixed movement sequence, and it sometimes improves on flawed demonstrations.
Skild describes S1 as the first robotics foundation model to show in-context learning on long-horizon tasks that were never seen during pretraining. Those tasks can run for up to 10 minutes, creating a demanding test of memory, control, and adaptation across many steps.
The model can perform unfamiliar tasks that include plant potting, pancake making, pour-over coffee brewing, and kit assembly. Each example points to the same core ability: S1 receives a visual example, interprets the intended result, and carries the task through without task-specific post-training.
An 11-Minute Path From Demonstration to Action
One plant-potting test shows how the process works in a real sequence. Soil, a pot, a watering can, and a plant arrived at 8:54 PM. Recording began at 9:16 PM, and an operator recorded one egocentric human video demonstration at 9:22 PM.
At 9:27 PM, S1 began executing the task. The time from demonstration to autonomous execution was 11 minutes, creating a clear timeline from a human example to robot action.
The test also highlights how S1 responds when the world refuses to stay fixed. The model can adjust when objects move mid-task, recover from errors, and substitute objects with matched affordance. Those abilities matter because a robot working through a multistep task cannot depend on every object remaining in exactly the same place.
S1’s approach treats a demonstration as an objective rather than a trajectory. That distinction gives the model room to recover when an action goes wrong or when an available object differs from the one shown in the video.
A More Than Sevenfold Gap In Multistep Tests
Skild’s tests on new, multistep tasks produced a sharp performance gap. S1 succeeded about 66% of the time at each step, while a similar AI system succeeded about 9% of the time at each step.
Skild characterized that difference as more than sevenfold. Because the tests involved new multistep tasks, the comparison centers on how each system handles sequences it did not encounter during pretraining.
- S1 uses a video demonstration as a prompt.
- The model executes tasks without updating its weights.
- It does not undergo task-specific post-training for each new task.
- It can handle tasks that run for up to 10 minutes.
- It can recover from errors and adapt when objects move.
- It can substitute objects with matched affordance.
Deepak Pathak, cofounder and CEO of Skild AI, framed the shift in direct terms: “Learning by experience, and not preprogramming, is the step change that has happened in robotics.” S1 puts that idea into a working system by connecting video demonstrations with autonomous task execution.
The combination of NVIDIA AI infrastructure, in-context learning, and long-horizon task performance gives Skild’s model a clear direction. S1 is not limited to repeating a stored routine; it is designed to take a new demonstration and act on its intended goal.
With Skild AI reaching a $100 million annual revenue run rate 10 months after its first commercial deployment, the model’s research results arrive alongside strong commercial momentum. The next stage for S1 will center on how far this demonstration-based learning can extend across unfamiliar tasks and changing environments.
Based on
