Dingkang Liang 介绍了团队的新研究 RoboScript，提出世界上下文学习（World In-Context Learning，WICL）。机器人在仿真世界中排练任务，从执行反馈中学习，并将可复用的知识积累在上下文里，整个过程无需更新模型参数。
作者报告，在每项任务仅提供一次演示、没有梯度更新的条件下，RoboScript 在 RoboTwin 2.0 全部 50 项任务上的平均成功率达到 83.4%。其中，干净环境的成功率为 83.6%，随机化环境为 83.2%，两种设置下表现接近。
这项工作的重点是让机器人利用自身尝试和执行反馈积累经验，将上下文学习延伸到与仿真世界的交互。Liang 将 WICL 定位为 VLA 训练的补充，探索预训练之后机器人继续适应任务的路径。原帖同时给出了论文与代码入口。
原帖留影
Can robots truly learn from experience? Recent advances in robotic in-context learning (ICL) have shown that robots can acquire new skills from a single demonstration, without updating model weights. But I've been thinking about another question: Can robots learn not only from demonstrations, but also from their own attempts, failures, and reflections? In our new work, RoboScript, we introduce World In-Context Learning (WICL). Instead of updating model parameters, robots rehearse tasks in a simulated world, learn from execution feedback, and accumulate reusable knowledge in their context. With just one demonstration per task and no gradient updates, RoboScript achieves 83.4% average success across all 50 RoboTwin 2.0 tasks, with nearly identical performance in clean (83.6%) and randomized (83.2%) environments. I don't see WICL as a replacement for VLA training, but rather as a complementary direction. Pretraining may determine where a robot starts, while learning from the world may determine how far it can grow. This is the direction we hope to continue exploring. 📄 Paper: https://arxiv.org/abs/2611.12885 💻 Code: https://github.com/ArcheOne-AI/RoboScript@Dinkang_Liang在 X 查看原帖发布于 2026/11/12 10:06（北京时间）