The model uses demonstration video as task context and applies a shared set of pretrained weights to execution. The research examines both whether a task was seen during pretraining and whether it requires a long sequence of actions, separating familiar imitation from harder forms of task and horizon generalization.
S1 is relevant to teams researching how robots can acquire new behaviors without collecting extensive new teleoperation data. The announcement highlights unseen tasks and roughly ten-minute horizons, while deeper training details are reserved for later publications. Public materials are available, but a downloadable model or priced self-service product is not established.
