S1 Robot Demo: Zero-Finetuning Execution of New Tasks from One Video

deepakpathak · x · 2026-08-28

A case shared by Joe Harris highlights the rapid adaptation of robot S1, requiring zero fine-tuning or post-training. By showing S1 a video demo of a new task, it infers the goal and translates it to its own body for execution. Skild moved from a potting demo to autonomous execution in just 11 minutes, with one prompt matching the efficacy of 380 post-training examples. S1 performed unseen workflows lasting up to 10 minutes and spanning dozens of steps, such as making pancakes, brewing coffee, and potting plants. While the average per-step success rate is 66% on internal benchmarks, this marks a shift where robots can be retasked rapidly via prompts.

Original post →

More from Embodied

Embodied channel →