Generated illustration for S1: In-Context Learning for Robotics — Paper 30 of 41Seedream 4.0 · GMI Studio

Illustration brief the model wrote

A single robot arm with a gripper stands before a table with a cup and a bottle. A smartphone on the table displays a video of a hand pouring from bottle to cup. The robot's gripper is reaching toward the cup, indicating it has learned the task from the video.

Provenance. Brief written by deepseek-ai/DeepSeek-V4-Flash via GMI Cloud on 2026-09-18 from the paper's title, primary link and the editor's notes; reviewed editorially. Graphic rendered by seedream-4-0-250828 through the GMI Studio request queue (request de1b23cd…), converted to PNG locally. Summaries can be wrong; the linked paper is the source of truth.

Atlas/Skild AI·Physical AI company·2026

S1: In-Context Learning for Robotics

Model tl;dr

S1 introduces in-context learning for robotics: a single video prompt specifies tasks, including unseen ones over ~10 minutes, with no weight updates and ~96% accuracy on seen tasks.

manipulationfoundation modelin-context learningvideo promptinglong-horizonzero-shot

What changed

Model

Before S1, robot manipulation required per-task data collection and fine-tuning. S1 removes these requirements by using in-context learning from one video prompt, enabling zero-shot execution of new tasks including long-horizon maneuvers up to 10 minutes without weight updates. It is already deployed with commercial partners, streamlining robot deployment.

So what, in plain English

Model

Instead of teaching a robot a new chore by showing it millions of examples and retraining it, you just show it a video once and it can copy you, even for jobs that last ten minutes. This means robots can learn new tasks on the fly, just from watching a short clip.

Contribution to industry

Model

S1 contributes a manipulation foundation model that eliminates per-task data collection and fine-tuning, reducing deployment cost and time. It provides a single video prompting interface for robot task specification and is already used by commercial partners for real-world applications.

Typed judgments

Jev

★★★★☆ manipulation · Physical AI 95%

Real hardware 58% · Open code / data 5% · Foundation-model based 84% · Survey / benchmark-only 4%

Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.

Editor's notes

Summary. A manipulation foundation model built for in-context learning: a single video prompt specifies a task, including unseen tasks over roughly 10-minute horizons, with no weight updates, reaching about 96% accuracy on seen tasks.

Industry. Removes per-task data collection and fine-tuning from robot deployment; the company reports S1 is already deployed with commercial partners.

Read the paper ↗Official technical research post ↗All 3 Skild AI papers →