Skild AI unveils S1 flagship robot foundation model

A company recently unveiled S1, a robot foundation model that enables robots to learn new complex tasks by watching a single video. The model uses in-context learning, meaning a human demonstration video can be added to the system's prompt and the robot can then perform the task, including long and intricate actions lasting up to 10 minutes. The approach matters because it contrasts with the current robotics practice of requiring lengthy post-training for each new task, which slows scaling. The foundation model achieves this by training on all four major types of robot training data together: teleoperation data, human videos, simulation, and data-capture gloves, with each source compensating for the limitations of the others.