RoboScript learns manipulation tasks in context from one demonstration
Original title: World In-Context Learning for Agentic Robot Manipulation
Authors report 83.4% mean success on all 50 RoboTwin 2.0 tasks without gradient updates, over 20 points above the strongest trained policy.
Researchers propose World In-Context Learning, in which a frozen model rehearses a demonstration in a simulator and stores feedback in context rather than weights. Their RoboScript system freezes that result as a Task Memory and deploys it on new scenes with data isolation meant to match a real robot. The authors say it generalizes across visual distributions, improves by reflection, scales with the foundation model, and transfers to a physical dual-arm robot. Code is at github.com/ArcheOne-AI/RoboScript.
Unconfirmed: Success rates and the comparison with trained policies are the authors' reported results.
- One demonstration per task and no weight updates.
- 83.4% mean success across clean and randomized RoboTwin 2.0 settings.
- Cross-embodiment memory adaptation is used on a physical dual-arm robot.
- Learned knowledge is frozen as a Task Memory before deployment.
- Products
- RoboScriptRoboTwin 2.0
- Organizations
- ArcheOne-AI
From the original article
1 min read at huggingface.co
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.
By Heng Fang, Zehui Guo, Yuehan Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.