A robot that learns from its mistakes: AGIBOT shows GO-3 AI model and heavy-duty G2 Max
At its Innovation Day in Shanghai, the Chinese robot maker presented a new robot brain, an industrial version of its G2 robot and its own metric for measuring robot work.
Photo: JJxFile / Wikimedia Commons, CC BY 4.0 (cropped and resized)
Video: YouTube · ShanghaiEye魔都眼 · youtube.com
Chinese robot maker AGIBOT unveiled a series of new products at its "Innovation Day 2026" in Shanghai on Thursday. At the centre is GO-3 Preview, a foundation AI model for robots that combines reasoning, predicting the outcome of actions and generating movements in a single framework. According to AGIBOT, it ranks first on three embodied-AI leaderboards – RoboDojo, RoboTwin and RoboColiseum. The model is to be fully open-sourced by the end of 2026. The video from Shanghai broadcaster ShanghaiEye shows the launch.
According to the company, a robot controlled by GO-3 identifies its own failures, adjusts its approach and stores what it has learned. In a test in a convenience store, the success rate for the full workflow rose from 28.6 to 85.7 percent within five days. Also new is the G2 Max for factories and logistics: it has 21 degrees of freedom, lifts up to 40 kilograms with both arms and works at heights between 0.1 and 2.5 metres. They are joined by a third hardware generation called AHA3 with liquid-cooled joints and the operating system Link-U OS 1.0.
With its new metric "Worken", AGIBOT wants to measure how much productive work robots actually deliver – calculated from the number of active robots, their autonomous working hours and their efficiency. "Worken is our attempt to develop a new way of answering that question," said co-founder and CTO Peng Zhihui. All performance figures come from the manufacturer; independent tests are pending.