According to CoinMeta, Skild AI has released a new generation of robot basic model, S1. The biggest change with this model is that to teach the robot new tasks, there is no need for re-tuning; it can learn directly by being shown a segment of human demonstration video. The tasks demonstrated by the official version lasted over 10 minutes and included tasks such as flipping pancakes, making espresso coffee by hand, and repotting plants, which had not been trained for before. S1 uses the demonstration video directly as a hint to guide the robot's actions. In the repotting experiment, it only took 11 minutes from the start of recording the human demonstration to the robot beginning to execute the task. Even if interrupted or made mistakes along the way, S1 will continue to adjust according to the current environment instead of mechanically repeating the actions from the video. Internal evaluations of Skild show that both models use the same 100,000 hours of data for pre-training. When faced with new tasks they have never seen before, S1 achieves an average success rate of 66% per step, while the traditional language instruction VLA only manages 9%. The latter would need to collect about 380 more task demonstrations and go through re-tuning to catch up to the level achieved by S1 after just one video viewing. S1 has already begun deployment to a small number of industrial customers, and the scope is expected to expand in the coming months.