Skild S1 reaches 66% on unseen robot tasks

Skild AI says its S1 robotics foundation model reaches 66% success on unseen manipulation tasks after a single video demonstration, versus 9% for language-prompted vision-language-action models trained on the same 100,000 hours of data, according to Skild’s technical announcement. The model handles tasks up to 10 minutes long without task-specific post-training, but the result is self-reported and leaves a meaningful failure rate; robotics teams should test video-conditioned policies against their own recovery and safety requirements, building on the archive’s physical-AI interface analysis.