At its 2026 Innovation Day on October 29, AgiBot co-founder, co-president and head of its embodied intelligence division Yao Maoqing unveiled GO-3 Preview, the company's latest embodied foundation model. It now ranks first on three embodied intelligence benchmarks — RoboDojo, RoboTwin and RoboColiseum — and will be fully open-sourced by the end of 2026.

A different number matters more than the leaderboards. Genie G2, running on AgiBot's embodied agent framework Genie Master, worked five consecutive days in a real convenience store, with its full-task success rate climbing from 28.6% on day one to a stable 85.7% — no human demonstrations, no engineers tuning it on site. AgiBot says this is the first time the industry has stepped outside simulation and the lab to validate Physical RSI, or embodied recursive self-improvement, in a live commercial setting.

Embodied Models Need Work, Not Just Training

Yao made a pointed argument on stage: embodied intelligence has to pursue full-pipeline, large-scale learning, and embodied models cannot simply be trained into existence — they have to be worked into existence.

That is aimed at the standard playbook of the past two years: pile up data in the lab, train a model, climb a leaderboard, and watch the capability collapse the moment it meets the real world. An embodied foundation model is meant to be a reusable base that transfers across tasks and across robot bodies, rather than a separate small model trained for each scenario. GO-3 Preview splits learning into three stages to deliver on that: pre-training builds general capability, post-training improves it through a mix of real and simulated interaction, and after deployment the model keeps reusing experience from actual work. Strung together, the robot keeps getting better while doing the job.

Pre-training: Packing What to Do, What Happens Next and How to Do It into One Model

In pre-training, GO-3 Preview uses WRAM (World Reasoning Action Model), an embodied-native architecture that jointly models task understanding, future prediction and action generation.

Those three jobs have typically been split across separate modules — parse the instruction, predict the state, plan the motion — passing data through interfaces. Unified modeling means prediction and action share a single understanding of the world, with no translation between two sets of representations.

Three designs make it work: CoAE, control-oriented visual encoding; SVP, single-step visual planning; and KASO, which screens actions for consistency with the predicted future. In plainer terms: first see the scene clearly, with visual encoding built to serve control decisions rather than human viewing; then predict what should come into view step by step, one intermediate goal at a time instead of planning far ahead in one shot; finally, when generating an action, check it against whether the future would actually unfold that way and discard the ones that do not match.

By AgiBot's account, this means the model not only understands what needs to be done, but can also predict what happens next and produce the specific actions to get there.

Post-training: Collective Experience from a Hundred Robots, with Imagination Filling the Gaps

Pre-training provides general capability; whether the robot can actually work comes down to post-training. For this stage AgiBot built Genie Evolver, a system that draws on a hundred physical robots alongside physics simulation and a world model.

Collective experience is the key idea here. Trial and error in the real world is expensive for a single robot — every fall and every collision costs time and money, and it can only follow one path at a time. Genie Evolver lets robots learn as a group: one real-world episode can be extrapolated into multiple future trajectories, effectively using imagination to multiply a single experience.

The division of labor is clear. Physical robots set the direction and supply real feedback, while simulation and the world model run through the large volume of trajectories that were never attempted on hardware. The world model's job here is predicting how the environment changes at the next step.

Those Five Days in the Convenience Store Were the Real Exam

The real test came after deployment.

Genie G2, running Genie Master, took up a post in a working convenience store. Without human demonstrations or tuning, it diagnosed its own failures, adjusted its strategy, and accumulated the experience. Full-task success was 28.6% on the first day and settled at 85.7% after five days.

Taken alone, 85.7% is not remarkable. What makes it different is how it got there: the number was not tuned in a lab, it rose on its own inside a live business with no human intervention. On the task of hanging small hooks, single-task success came in 17.7 percentage points above conventional fine-tuning. Conventional fine-tuning relies on humans labeling data and designing training tasks; here the robot gathered it on the job.

AgiBot calls this Physical RSI, or embodied recursive self-improvement: the operational data the robot generates is fed back into the model itself.

From ViLLA to GO-3, Three Generations Betting on the Same Thing

AgiBot has been on this path for three generations. GO-1 introduced the ViLLA architecture, GO-2 pioneered an action chain-of-thought, and GO-3 is about full-pipeline, large-scale learning.

All three are answering the same question: can physical AI keep scaling on data and feedback the way large language models do? Language models scale because the text corpus is already sitting on the internet. Robots have no such ready-made corpus, so they have to produce it themselves while working.

GO-3 Preview is slated for full open-sourcing by the end of 2026. Whether experience generated by other companies' robots can also feed into this loop is the next thing to be proven.