Found first: a primary source the press has not covered yet.
A team of researchers has published Zero-WAM, a causal video-action model that treats human demonstration videos as task specifications for robot manipulation, without any per-task fine-tuning. On seven unseen tasks in the RoboTwin 2.0 simulation benchmark, it achieves 46.95% average success, a 29.50 percentage point improvement over the strongest video-action baseline. The paper appeared on arXiv on 27 August 2026.
What the source says
Jiaming Zhou, Qihang Zhang, and ten co-authors built HumanGen, an automatic pipeline that converts robot trajectories into semantically matched human demonstration videos, producing 74.2K human-robot in-context learning pairs across 8.6K distinct tasks. The core training objective, in-context future chunk prediction (IFP), forces the model to rely on the provided human video rather than fall back on its own robot-action history. On RoboTwin 2.0, Zero-WAM scores 46.95% averaged across seven unseen manipulation tasks, against 17.45% for LingBot-VA and 10.98% for WAN-Action. Real-world tests span multi-object placement (53.3% success vs. 43.3% for the baseline), sequential three-object manipulation (33.3% vs. 10.0%), and table-leg insertion (16.7% vs. 0.0%).
Why it matters
In-context learning lets language models adapt to new tasks at inference time from a handful of examples, and applying that mechanism to physical manipulation has been a recurring research target. Zero-WAM's result is that human demo videos carry enough task information to substitute for per-task training, at least on the tasks tested. The 29.50 percentage point gap on unseen tasks is measured against the best existing video-action baseline on a public benchmark, making it a verifiable claim. On the insertion task, where the baseline scored 0%, Zero-WAM reached 16.7%, a harder test of generalization than aggregate success rate alone.