Found first: a primary source the press has not covered yet.
Researchers at Huazhong University of Science and Technology and ArcheOne AI have shown that a robot policy with frozen weights can outperform trained action policies on a large manipulation benchmark. Their system, RoboScript, achieves 83.4% mean success across 50 tasks on RoboTwin 2.0 using one demonstration per task and no gradient updates.
What the source says
RoboScript implements World In-Context Learning (WICL): given one demonstration per task, a frozen foundation model infers task steps and object roles from context rather than from weight updates. The evaluation covers 50 tasks on RoboTwin 2.0 in both clean and randomized visual settings, yielding 83.4% mean success. That exceeds the strongest trained action policy by over 20 percentage points. The authors report that the system's representations are semantic, encoding what objects do in a task rather than what they look like, which allows experience from one visual environment to transfer to another. The paper also reports performance improvements through reflection, scaling with the foundation model, and transfer to a physical dual-arm robot via cross-embodiment memory adaptation.
Why it matters
The standard assumption in robot learning is that competitive task performance requires gradient updates on task-specific data. The authors report this is the first demonstration of an in-context approach exceeding trained policies across a 50-task suite without touching model weights. The semantic rather than visual representation is the mechanism: one demonstration generalizes across visual distributions because the model is not fitting to appearance. For practitioners, this reduces the data and compute cost of deploying a new manipulation task. The result also implies that scaling the underlying foundation model improves robot task performance directly, without any task-specific fine-tuning step.