Skild AI Publishes Manipulation Results for S1 - Show It One Video and It Runs Tasks It Never Trained On, Company Says

Robotics company Skild AI published manipulation results for its foundation model S1 on its blog¹. The company describes S1 as designed from the outset to learn in context, and says a video is all it needs in order to attempt a task, whether that task is a quick one or a long one and whether or not it appeared in training¹.

The distinguishing choice is how a task gets communicated: a video demonstration takes the place of a written or spoken instruction¹. Because pre-training spans many different tasks, the company argues, the model has to work out what the demonstrator is trying to do, which it says leaves one set of weights able to handle tasks absent from training, with no additional tuning¹. It states plainly that there is no fine-tuning and no post-training involved¹. By its own account of the earlier approach, every new task meant collecting hours of data and tuning a specialist policy for it, and for tasks of moderate complexity the volume of post-training data had to run from tens into hundreds of hours recorded under the conditions where the robot would actually work¹.

Skild AI presents this as a first: by its account, no robotics foundation model before S1 has demonstrated in-context learning on tasks that run as long as ten minutes and were absent from pre-training¹. These are the company’s own claims, not third-party verification. It says a future post will cover how S1 is trained¹.

Sources

1. Introducing S1: In-Context Learning for Robotics - Skild AI official blog