Skild S1 robot model learns new jobs from one video

Skild introduced S1 on August 18, with the announcement appearing in The Rundown's September 3 newsletter. The core idea is to apply a heavily pretrained foundation model (100,000 hours of pretraining) to new extended tasks by supplying a video as an input prompt, rather than updating weights. This positions S1 as a test of rapid adaptation for multi-step jobs that can take up to ten minutes.

The company's controlled comparison gives scale to the potential savings: one video prompt delivered performance that would otherwise require roughly 380 teleoperated demonstrations (50–100 hours). However, Skild's metric averages success across steps and includes human recovery interventions, meaning full-job completion rates without assistance remain unclear. A practical workflow may involve demonstrating a task, assessing performance, then collecting targeted additional data where reliability falls short.

Generalist offers a related example, reporting 59% average success across ten simple, short tasks from one 3–12 second demonstration without weight updates, and 83% after ten gradient steps with five minutes of task data. Skild's larger demonstration collection also exceeded its single-video score. This suggests hybrid approaches — initial video prompting followed by focused training — could balance speed and reliability, though how well these methods generalize across new environments and hardware remains an open question.