Skild AI Unveils S1 Robot Model That Learns From One Video

Skild AI has introduced S1, a robotics foundation model designed to bring a new form of in-context learning to physical machines. Instead of requiring engineers to collect large amounts of new training data and retrain a robot for every task, S1 can watch a video demonstration and use that example to perform the task. Skild says the system can handle tasks lasting up to 10 minutes.

The key innovation is that the video itself becomes the instruction. S1 analyzes the demonstration to understand the goal, the sequence of actions and the progress of the task, then translates that information into robot actions. Skild says this allows the model to perform tasks that were not present in its pre-training data, potentially reducing the amount of task-specific engineering traditionally required in robotics.

Skild demonstrated S1 on several long-horizon tasks, including repotting a plant, preparing pancakes, making pour-over coffee and assembling a kit. These tasks involve multiple steps and were described by the company as unseen during pre-training. The demonstrations are particularly notable because the robot is not simply repeating a short movement; it must maintain context across a much longer sequence of physical actions.

The technology could have major implications for factories, warehouses and other environments where robots frequently need to learn new procedures. Skild’s broader goal is to develop an “omni-bodied” intelligence layer that can control different types of robots rather than building a separate AI system for every machine. The company says S1 is already being used with commercial partners.

There is still an important distinction between a company demonstration and independently verified performance. Skild’s reported results show the potential of one-shot learning for robotics, but broader testing will be needed to determine how reliably S1 performs across different robots, environments and repeated trials. If the approach proves robust, however, it could represent an important step toward robots that can learn new physical skills almost as easily as humans learn by watching someone demonstrate a task.