Daily papers

Wednesday, 11 November 2026 · 5 min read

World In-Context Learning enables robots to learn manipulation tasks from one demo with simulation feedback

World In-Context Learning (WICL) uses a frozen model that rehearses a task in simulation after one demonstration, storing feedback in context rather than weights, achieving 83.4% success on 50 RoboTwin 2.0 tasks without gradient updates and surpassing trained policies by 20 points.

Based on the paper World In-Context Learning for Agentic Robot Manipulation by Heng Fang, Zehui Guo, Yuehan Zhang, Mingyang Du and 2 others

Challenges in robot manipulation learning and limitations of trained policies

Robotic manipulation tasks require learning action policies that map observations to actions. Traditional approaches train these policies end-to-end, embedding task knowledge into network weights. This tightly couples the learned skills to the training distribution, limiting generalization to new scenes or embodiments.

In-context learning (ICL) offers an alternative by conditioning a frozen model on demonstrations at inference time, avoiding weight updates. However, prior ICL methods for manipulation have underperformed trained policies because demonstrations from other scenes lack feedback about the current world state, making it hard to adapt actions appropriately.

The key insight is that what an agent learns in context is semantic: it records task steps and object roles rather than specific visual details. This suggests that experience gathered in one world can transfer to another if feedback is incorporated properly, motivating a new paradigm that integrates world feedback into the context rather than the weights.

Sign in to read the whole paper

You have read the opening. The rest is free with a Bezen account, and no card is needed.

In the rest of this paper

  • How the method works, step by step, with a flow chart
  • How it was tested, what was found, and where it falls short
World In-Context Learning enables robots to learn manipulation tasks from one demo with simulation feedback · Bezen