# World In-Context Learning for Agentic Robot Manipulation

**Authors:** Heng Fang, Zehui Guo, Yuehan Zhang, Mingyang Du, Dingkang Liang, Xiang Bai  
**Lab:** ArcheOne-AI · **Published:** 9 Nov 2026 · **Upvotes:** 64 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.12885) · [PDF](https://arxiv.org/pdf/2611.12885) · [Code](https://github.com/ArcheOne-AI/RoboScript) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.12885.html)

## Curator's notes

**TL;DR:** Robots learn manipulation from a single demo by rehearsing in simulation, writing feedback to context, achieving 83.4% success without gradient updates.

- **What's new:** Proposes World In-Context Learning where a frozen model rehearses tasks in a simulated world and stores feedback in context, surpassing trained policies with one demonstration.
- **Key result:** RoboScript reaches 83.4% mean success on all 50 RoboTwin 2.0 tasks, exceeding the strongest trained action policy by over 20 points.
- **Why it matters:** Allows rapid skill acquisition for robots without costly retraining, leveraging simulation to adapt to new scenes.

## Key numbers

- **83.4%**: RoboScript reaches 83.4% mean success across clean and randomized settings with one demonstration per task and…
- **50**: On all 50 tasks of RoboTwin 2.0
- **20**: exceeding the strongest trained action policy by over 20 points

## Abstract

Trained action policies write manipulation skill into network weights, tying what is learned to the distribution it is learned from. In-context learning (ICL) avoids this coupling but has not matched trained policies, because a demonstration recorded in another scene carries no feedback from the world the robot currently faces. We observe that what an agent learns in context is semantic, recording task steps and object roles rather than visual appearance, so experience gathered in one world transfers to another. Based on this insight, we propose World In-Context Learning (WICL), a paradigm in which a frozen model starts from a demonstration, rehearses the task in a world such as a simulator, and writes the feedback into its context rather than its weights. We build RoboScript, a modular infrastructure that implements WICL for agentic manipulation. From a single demonstration, RoboScript learns each task through rehearsal in simulation, freezes the learned knowledge as a Task Memory, and deploys the frozen memory on new scenes with data isolation that ensures every deployment input is available on a real robot. On all 50 tasks of RoboTwin 2.0, RoboScript reaches 83.4% mean success across clean and randomized settings with one demonstration per task and no gradient updates, exceeding the strongest trained action policy by over 20 points. The method generalizes across visual distributions, improves through reflection, scales with the foundation model, and transfers to a physical dual-arm robot through cross-embodiment memory adaptation. With a world in the loop, a robot can learn a manipulation task in its context as well as it learns one in its weights. All code is available at https://github.com/ArcheOne-AI/RoboScript

## What you could build

1. Create a demo-to-simulation pipeline that lets a robot learn new pick-and-place tasks from a single video.
2. Develop a cross-embodiment memory adapter to transfer learned Task Memories between different robot arms.
3. Build an interactive UI for users to provide one-shot demonstrations and instantly deploy them on a real robot.

## Quiz yourself

What mean success rate did RoboScript achieve on the 50 RoboTwin 2.0 tasks with one demonstration per task?

**Answer:** It achieved an 83.4% mean success rate, which is more than 20 points higher than the best trained action policy.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.