During testing, the model successfully performed novel, long-horizon tasks lasting up to 10 minutes, such as potting a plant and cooking pancakes, without any prior training on those specific activities.
This development addresses a major bottleneck in AI infrastructure where robots typically remain rigid, specialized tools.
By enabling a "prompting" workflow for physical actions, Skild AI reported that S1 can be deployed for a new task in approximately 11 minutes, compared to the hours or days required for conventional manual data collection.
The company's research indicates that this approach is significantly more effective than language-based instructions for novel skills; in a study of unseen tasks, the in-context model achieved a 66% success rate, whereas models relying on language prompts achieved only 9% success.
Built on NVIDIA accelerated computing, S1 also demonstrates emergent common-sense reasoning and physical robustness.
According to Skild AI, the model can recover from unexpected errors, such as dropping an item, and can adapt to environmental changes like shifting light or moving objects that were not present in the video demo.
This capability is intended to help general-purpose robots scale across diverse commercial environments where conditions change daily and manual reprogramming is impractical.