Chinese robot maker AGIBOT unveils GO-3, an AI brain meant to keep learning on the job
The company says the model tops three robot tests and will be released free by the end of 2026.
AGIBOT, a Shanghai company that builds humanoid and other robots, on Tuesday unveiled GO-3 Preview, the AI software that decides how its robots see and move. The company said it will make the model free and open by the end of 2026.
GO-3 is an embodied foundation model: one large AI model trained on many tasks that controls a physical robot, rather than just producing text. AGIBOT said it ranks first on three public robot tests, RoboDojo, RoboTwin and RoboColiseum.
How robots usually learn
Most robots today learn to handle objects by copying people. In the usual method, called teleoperation, a worker steers the robot by remote control, with a joystick, a virtual reality headset or a second robot arm moved by hand, while the robot records what it saw and how it moved. The robot's AI is then trained to repeat those moves, many times over for each kind of task.
Those examples are high quality but expensive and slow to collect, according to a 2026 research paper presented with CVPR, a leading computer vision conference. Two cheaper sources, computer simulations and videos of people working, are easy to gather at large scale, but they don't match a real robot's body or camera view, the paper said. And once a robot is at work, fixing a problem it has never seen has usually meant human engineers rewriting code, collecting more data or adjusting the robot by hand, AGIBOT said.
What AGIBOT says is different
GO-3 also starts from videos of people's hands at work and from recorded robot motions, so that part is not new. AGIBOT said the model predicts what a scene will look like next, then picks moves to get there. It then practices on a fleet of 100 real robots, in physics simulators and in world models, AI that imagines how a scene could play out, and what one robot learns is shared with the rest.
The bigger change, AGIBOT said, comes after a robot is put to work. Software called Genie Master logs each success and failure on the job, saves the lessons each night as long-term memory, and feeds the records back in to retrain how the robot moves, so the robot improves from its own work rather than from new human demonstrations. AGIBOT said a Genie G2 robot in a real convenience store kept improving at hard tasks without human demonstrations or tuning.
The rankings and results are AGIBOT's claims and have not been checked independently. The company said its 20,000th robot came off the line in September.
Sources: AGIBOT news release via PR Newswire, Nov. 3, 2026; "Teleoperation, Simulation, or Human Video? Data Utilization Law for Robot Manipulation," CVPR 2026 paper (Computer Vision Foundation)