For years, teaching a robot a new task has meant collecting hundreds of demonstrations and retraining. Skild AI wants to end that. On 25 August the company introduced S1, a robot foundation model that learns from a single human video and acts on it straight away, without fine-tuning or any task-specific training afterward.
The idea borrows from how large language models handle instructions. Show a language model a couple of examples in the prompt and it adapts on the spot. S1 does the same with motion. Give it one video of a job, whether the model has seen anything like it before or not, whether it takes ten seconds or ten minutes, and it attempts the task directly.
The numbers
Skild reports that S1 succeeded on 66% of unseen tasks in its internal benchmarks. A vision-language-action model prompted with words instead of video, trained on the same 100,000 hours of data, managed 9%. The company also estimates that a single video prompt buys performance roughly equal to 380 task-specific demonstrations collected the old way. If that holds up outside Skild's own lab, it changes the economics of getting a robot to do something new.
The demonstrations were chosen to make the point. Skild showed the robot potting a plant, cooking pancakes, making pour-over coffee and assembling a kit, none of which appeared in its pretraining. These are fiddly, multi-step jobs of the kind that usually defeat general-purpose robots.
Worth a note of caution
There is a catch, and Skild is upfront about it. There are no released weights, no API and no paper. Everything public rests on the company's own figures and a set of curated clips. That is not unusual for a young robotics firm guarding its lead, but it means the results cannot yet be reproduced by anyone outside. Skild raised $1.4 billion earlier this year at a valuation above $14 billion, so the incentive to impress is considerable.
Even discounted, the direction is striking. The robotics field has spent the past year chasing a general model that transfers across tasks, the way a language model transfers across topics. Chinese firms have poured money into humanoid hardware, from XPeng's $900 million robot unit to the record-setting Tiangong humanoid. Hardware, though, has never been the hard part. The brain is. S1 is a claim that the brain is starting to catch up, and one that independent researchers will want to test the moment Skild lets them.
Commentarii · 0