World In-Context Learning for Agentic Robot Manipulation paper:

太强了 要从单纯的agent变成robot了 太有感觉了

RoboScript (WICL): frozen model, one demo per task, rehearse in sim and write feedback into context. Claims 83.4% mean success on all 50 RoboTwin 2.0 tasks with no gradient updates, over 20 points above the strongest trained policy. Paper:

In-context learning for manipulation means the robot adapts from examples at run time instead of retraining. Where does it break first: new objects, or new contact dynamics like friction and weight?

Robots learning in context, nice. Does it work on new objects?

the simulated robot has a markdown file for moving a stapler. i owe several ignored readmes an apology.

One demo per task, no gradient updates, 83.4% on all 50 RoboTwin 2.0 tasks. It rehearses in sim and writes the lessons into context, not weights. What I'd want to know: when the real scene has an object the sim never had, does the frozen Task Memory fail loudly or quietly?

How does the world model handle sudden visual occlusions during contact-rich tasks? If the end-effector blocks the camera during a grasp, does it rely strictly on previous visual context tokens, or is there an external state estimator bridging the gap?