AgiBot trains a model from scratch on 30,000 hours of real robots: towel folding appears only at the largest scale
AgiBot publishes GE-Act 2.0, trained from scratch on data from real robots. Taking the hours of data from 300 to 30,000, success on never-seen tasks rises from 17.1% to 44.1%. The household skills — folding towels, arranging flowers — appear only at the largest scale. It remains research.
⚠️ How to read this piece. This is a research result, not a product. The source is the technical report by AgiBot's team deposited on arXiv on September 4, 2026: all figures are declared by the authors and have not been independently verified by any third party. We report them as such, with the source next to them.
In brief
- AgiBot's research team has published GE-Act 2.0, a model that first imagines the future state of the scene and then derives the robot's actions from it, instead of simply repeating trajectories guided remotely by a human operator. (source: arXiv:2609.05588, September 4, 2026)
- The declared methodological novelty is that all components start from scratch, trained on robotic manipulation data, without inheriting a video generator already trained on web material. According to the authors, almost all previous work does the opposite, leaving unverified whether manipulation-specific training really scales. (source: the technical report)
- The scaling curve is the central result. Taking the data from 300 to 30,000 hours, success on never-before-seen tasks — no dedicated training, new scenes, backgrounds, lighting and objects — goes from 17.1% to 44.1% on one robot body and from 13.4% to 31.1% on a second. (source: the technical report)
- The test bench consists of 100 real tasks across 20 skill groups, with two different robot bodies. The improvements cover 19 groups out of 20 in the first case and 18 out of 20 in the second. (source: the technical report)
- Data collected on one robot improve a different robot: the second body accounts for less than 2% of the training material, yet gains 17.7 points, in percentage terms, as the shared corpus grows. (source: the technical report)
- How much data exists for a single skill strongly correlates with success in new scenes. In plain terms: there are no shortcuts, you need data of that gesture. (source: the technical report)
- On language understanding, the model correctly grounds references to object, color, shape and position in at least 90% of trials, and follows an explicit instruction even when it contradicts a movement already under way. (source: the technical report)
Why it matters to us: the household angle
The skills the authors indicate as emerging only at the largest scale — that is, completely absent in the model trained on 300 hours — are, as it happens, household gestures: folding towels, stacking paper cups, taking the cap off a pen, arranging flowers. (source: the technical report)
Our reading. It is the same direction as the race for household data we have already written about in Who collects the data inside homes (in Italian): the bottleneck of the home robot is no longer the body, it is the repertoire of gestures. And the repertoire is bought in hours of data. Here, for the first time, there is a public measure of how much it costs: to make towel folding appear, it took one hundred times the starting hours.
A detail that says a lot about the method: in the training material the authors also include failed trajectories, treated as useful information rather than as noise to be thrown away. (source: the technical report)
What it does not say
- 44.1% is not household reliability. On 100 tasks, the best model fails more than half the time. A robot that gets one attempt in two wrong in the living room is not a product: it is an experiment.
- It is not a product on sale. There is no AgiBot home robot that can be bought in Italy, nor a price, nor a date.
- The two robot bodies tested are lab and work platforms, not robots certified for European home use.
- The curve did not flatten between 5,000 and 30,000 hours, and this is the story behind the story. If the next leap requires another order of magnitude of data collected in the real world, the real question becomes who can afford to collect it — a structural Chinese advantage we have already written about in connection with shipments in the first half of 2026 (in Italian). This is our reading, not a figure from the report.
Sources
- arXiv (authors’ preprint) primary source
- AgiBot — GE-Act 2.0 (authors’ project page) primary source
- Hugging Face unclassified
- Pandaily press
Next to each source we state what kind of source it is: “primary source” means the manufacturer, an official document or the author of the data; “press” a news report; “retailer” a distributor or a store, reliable about its own offer — price and availability in Italy — but not about the product’s features; “analysis” a professional firm or a research center that signs its own reading; “Chinese state media” an outlet controlled by the state, whose content should be weighed as such; “blog/forum” a non-editorial source; “unclassified” a source we have not assessed yet: the link stays, the judgment does not.