RoboScript Hits 83.4% on RoboTwin 2.0 From One Demo Per Task
TL;DR
- A frozen model reached 83.4% mean success on all 50 RoboTwin 2.0 tasks with one demonstration per task and no weight updates.
- The reported gap over the strongest trained action policy baseline is more than 20 points, though the baseline model is not named.
- The system freezes in-context experience as a Task Memory and claims transfer to a physical dual-arm robot via cross-embodiment adaptation.
A frozen foundation-model policy reached 83.4% mean success across all 50 tasks of the RoboTwin 2.0 benchmark with one demonstration per task and no gradient updates, "exceeding the strongest trained action policy by over 20 points," according to a new arXiv preprint from Heng Fang and colleagues.
The authors call the paradigm World In-Context Learning, implemented as a modular stack named RoboScript. From a single demonstration, RoboScript "rehearses the task in a world such as a simulator, and writes the feedback into its context rather than its weights," before freezing what it learns as a Task Memory and deploying that memory on new scenes.
"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," the paper concludes.
The 83.4% figure covers clean and randomized settings together; the abstract names neither the specific trained-policy comparator nor any per-task numbers. The method "transfers to a physical dual-arm robot through cross-embodiment memory adaptation," though no real-robot success rate is reported in the abstract. Code is published at the authors' repository.
Originally reported by paper
Read the original article →Original headline: World In-Context Learning: Frozen Robot Policy Clears 83.4% on 50 Tasks With One Demo, Beats Trained Policies by 20pp