Found first: a primary source the press has not covered yet.
Researchers at Huazhong University of Science and Technology and ArcheOne AI have introduced World In-Context Learning (WICL), a paradigm that lets a frozen robot policy adapt to new manipulation tasks from a single demonstration, with no gradient updates at any stage. Their system, RoboScript, achieves 83.4% mean success across 50 tasks on the RoboTwin 2.0 benchmark, exceeding the strongest trained action policy baseline by over 20 percentage points.
What the source says
The paper, submitted November 9 2026 by Heng Fang, Zehui Guo, Yuehan Zhang, Mingyang Du, Dingkang Liang, and Xiang Bai, describes WICL as a loop in which the frozen model starts from one demonstration, rehearses the task in a simulator, and writes feedback into its context rather than its weights. RoboTwin 2.0 covers 50 tasks evaluated in both clean and randomized settings. The model learns task steps and object roles rather than visual appearance, which the authors say enables transfer across visual distributions. The paper also reports that performance scales with foundation model size and that the approach transfers to physical dual-arm robots via cross-embodiment adaptation.
Why it matters
Matching weight-trained robot policies using in-context examples alone, across a 50-task benchmark, has not been demonstrated before at this margin. WICL does it without touching model weights, which means adding a new task requires one demonstration rather than a training run. For teams deploying generalist manipulation policies, that changes the cost structure of task acquisition. The transfer to physical dual-arm hardware suggests the benchmark result is not purely a simulation artifact, though the paper does not report physical success rates.