Skild AI S1

2026-08-18 · Publisher : Skild AI · Domain : Robotics

Skild AI unveils S1, its flagship foundation model built from the ground up for in-context learning: it performs an unseen task, up to ten minutes long, from a single video and without post-training.

### Background On August 18, 2026, Skild AI introduced S1, its flagship robot foundation model, built from the ground up as an in-context learner. Show it a video of a task, short or long, seen or unseen, and it executes it. ### Specifications - **In-context learning**: a single video serves as the prompt; no post-training or extra data collection is needed. - **Scope**: unseen tasks over horizons of up to ten minutes, per Skild AI. - **Thesis**: Skild argues robotics has been stuck in the "BERT era", where every new application requires hours of data and a fine-tuned specialist. For the company, the main point of pretraining is to enable the same leap as from BERT to GPT-3: learning from one or a few examples. - **Milestones**: an in-context learner for locomotion released a year earlier, then first manipulation results in early 2026 before S1. - **Announced follow-up**: this post is the first in a series detailing how S1 is built. ### Why it matters If a robot can learn a task by watching one video, deployment costs collapse: no more tens of hours of demonstrations per task and per customer. It is robotics' equivalent of the prompt. ### Key takeaway Skild AI joins Generalist AI (GEN-1.5, released the next day) in the race for one-shot learning, eight months after its $1.4 billion raise.

Tags : Apprentissage en contexte, Modèle de fondation, Skild AI, Manipulation