Generalist AI's GEN-0 claims robotics' first scaling laws, trained on 270,000 hours of real-world manipulation data

Generalist AI, a robotics start-up with offices in the San Francisco Bay Area and Boston, introduced GEN-0 on November 4, 2025, describing it as a new class of embodied foundation models trained directly on raw physical interaction. The company's central claim is that robot capability now scales with data and compute in a predictable way, as it has in language models, and that robots below a certain size cannot absorb the data at all.

The numbers in the research post are large by the standards of the field. Generalist said GEN-0 was pretrained on its in-house dataset of more than 270,000 hours of real-world manipulation data gathered in thousands of homes, warehouses and workplaces, and that its operations were adding more than 10,000 hours a week. It reported training runs at 1 billion, 6 billion and 7 billion parameters and said it had since scaled GEN-0 beyond 10 billion parameters.

The headline scientific claim is a phase transition at 7 billion parameters. According to Generalist, 1 billion parameter models showed ossification under data overload, meaning their weights stopped absorbing new information during pretraining, while 6 billion parameter models began to benefit and showed strong multi-task ability. Models of 7 billion parameters and above internalised the pretraining data and transferred it to new tasks with only a few thousand steps of post-training, the company said, adding that it believed this was the first time ossification had been observed in robotics.

The second claim is a scaling law. Generalist said downstream validation error after fine-tuning followed a power law in the size of the pretraining dataset, of the form L(D) = (D_c / D)^α, across 16 task sets that included Lego building, fast food packing and open-ended tasks. In blind A/B tests on real robots, more pretraining data raised task success for models post-trained on only 5.6 hours of task data, equal to 1% of the 550-plus hours available, and the best results, up to 99% in certain cases, came from pairing full pretraining with all task data. The company said the pretraining and post-training sets were collected by different people in different environments.

Generalist also introduced what it calls Harmonic Reasoning, a training method in which the model handles asynchronous, continuous-time streams of sensing and acting tokens so that it can think and act at the same moment. The post presented this as an alternative to the two-system designs used elsewhere, citing Figure AI's Helix as an example, and to inference-time guidance. The company said GEN-0 had run on robots with 6 and 7 degrees of freedom and on semi-humanoid robots with 16 or more, and showed a long-horizon task in which a robot assembles a camera kit in a single stream without explicit subtasks.

The infrastructure disclosures are unusually specific. Generalist said it built custom hardware and dataloaders, laid new dedicated internet lines to collection sites, signed multi-cloud contracts, ran data processing on the order of 10,000 CPU cores and compressed dozens of petabytes, giving it the ability to absorb 6.85 years of manipulation experience per day of training. That is about 60,000 hours a day, which means the whole 270,000-hour corpus passes through training in roughly four and a half days, by MEKAWIRE's calculation. The post did not name its GPU supplier, the model architecture in detail, or its data foundry partners, which it labelled only as Partners A, B and C.

GEN-0 was not released as open weights or a public API in the post, and no technical report or preprint accompanied it. Generalist's earlier public material was a research preview in June 2025. The company's about page says its team previously built the PaLM-E and RT-2 embodied models and came from OpenAI, Boston Dynamics and Google DeepMind.

The comparison with rivals rests on disclosed scale. Physical Intelligence's π0, described in an October 2024 arXiv preprint, has 3.3 billion parameters and was pretrained on more than 10,000 hours of robot data; its π0.5 preprint of April 2025 used about 400 hours of mobile manipulator data from about 100 homes within a wider co-training mix. NVIDIA's GR00T N1, released in March 2025 with open weights, has 2.2 billion parameters and drew on 88 hours of real teleoperation, 827 hours of generated video and simulated data equal to 6,500 hours, its preprint says. Google DeepMind's Gemini Robotics, built on Gemini 2.0 and updated to version 1.5 in September 2025, has not disclosed parameter counts or dataset sizes. On disclosed figures, GEN-0's corpus is up to 27 times π0's.

The caveats are material. Every result comes from Generalist's own evaluations, the scaling plots are reported without trial counts, and the 99% peak applies to unnamed tasks. The 7 billion parameter threshold was observed on Generalist's data mix and training recipe and may not carry over to other labs. Comparing hours across companies is also imperfect, because Generalist did not break its corpus down by robot data and human-worn device data.

Later events show where GEN-0 led. In March 2026 Generalist ran a live GEN-0 demonstration at NVIDIA GTC on a new Universal Robots mobile platform with UR7e arms on a MiR base, which it said it prepared in a few days using no data from the exhibition hall. On April 2, 2026 it released GEN-1, pretrained on half a million hours of physical interaction data with no robot data, which it says lifts average task success to 99% from GEN-0's 64%. On June 4, 2026 it announced $400 million led by Radical Ventures, with NVIDIA among existing investors.