Skild's S1: a robot learns a 10-minute task from one video demo
In-context learning finally reaches robotics: one demo replaces ~380 episodes, 11 minutes to autonomous — but access is early-invite only for now.
Copy markdown
One video, a 10-minute task, no fine-tuning
S1 watches a single human demo and then executes tasks up to 10 minutes long that it has never seen, with zero task-specific training. Skild says one demonstration is worth roughly 380 post-training episodes — the in-context-learning trick that made LLMs general, now arriving for robots.
The numbers that make it a milestone
On out-of-distribution tasks S1 hits 66% success at 100k training hours and beats language-prompted policies by about 7x on unseen work. It generalizes from the demo instead of leaning on a fixed, pre-trained skill library.
Eleven minutes from demo to doing it
Skild clocks about 33 minutes from receiving materials to autonomous execution — and just 11 minutes from recording the demonstration. Teaching a robot starts to look like recording a Loom, not writing and tuning a policy.
The catch: you can't build on it yet
There's no public API and no open weights — S1 is running with commercial partners behind an early-access sign-up. File this under 'watch,' not 'ship': it signals a new buildable category coming, not a weekend project today.
Elsewhere in your stack
GitHub Copilot CLI v1.0.81 (Aug 27) adds a plugins dashboard and better session recovery. And Cursor's cloud agents are now drivable straight from the Vercel AI SDK via @ai-sdk/harness-cursor's HarnessAgent interface.