FreshSynapse

AI trends & daily updates, in plain English

AI Models

AGIBOT's GO-3 Preview Tops Three Robot Benchmarks, Promised Open-Source by Year-End

By FreshSynapse Editorial · · 3 min read

A robotic arm and a human hand reaching toward each other

Photo: Hippopx (public domain).

What happened. Chinese robotics company AGIBOT has unveiled GO-3 Preview, an embodied AI foundation model designed to help robots learn continuously — and says it ranks first on three embodied AI benchmarks, with a full open-source release promised by the end of 2026.

The model was announced at AGIBOT Innovation Day in Shanghai on 29 October 2026. It is built on what the company calls WRAM, the World Reasoning Action Model, an architecture that unifies three jobs: understanding a task, predicting what will happen next, and generating the physical actions to carry it out. The idea is a robot that can take an instruction, anticipate outcomes, and translate decisions into movement — then keep improving through real-world experience.

The benchmark claims. According to AGIBOT, GO-3 Preview secured first place on RoboDojo, RoboTwin and RoboColiseum. Reporting on the announcement adds detail: first in task success rate and process score on RoboDojo, the highest overall score on RoboTwin, and first overall on RoboColiseum. The company also says that, paired with its Genie Evolver post-training system — which combines around 100 real robots with physics simulation and world models — a robot’s task effectiveness rose from about 28% to stabilise at around 85% after five days of operation. These are company-reported figures from its own event, not independently reproduced results, and should be read that way.

Why the caveat matters. One of the three benchmarks, RoboColiseum, is built on AGIBOT’s own Genie Sim benchmark, so it is not independent of the company. Independent confirmation of the other two results on public leaderboards was partial at the time of reporting: RoboDojo’s public leaderboard listed a GO-3 entry in first place, while the RoboTwin result could not be confirmed on the public pages checked. The open-source promise is therefore the most important part of the announcement — if the full model is released as planned, outside groups will be able to test the claims on their own tasks.

The same event also launched Lingqu OS 1.0, an operating system built for embodied intelligence that AGIBOT says will also be fully open-sourced by the end of 2026 and donated to the OpenAtom Foundation, and a new “Worken” metric — effective intelligent work hours, defined as active robots multiplied by autonomous working hours multiplied by an output efficiency ratio — as a proposed way to measure what deployed robots actually produce.

For the wider field, GO-3 Preview is a marker of where embodied AI is heading: away from robots that repeat one trained motion, and toward models that generalise across tasks and keep learning after deployment. Whether AGIBOT’s version truly leads the field will be settled not by its event presentation, but by the open release it has promised and by what independent researchers find when they run it.

Sources

  • AGIBOT
  • GO-3
  • robotics
  • embodied AI
  • humanoid robots
  • benchmarks
  • open source

Related articles