Zhiyuan Releases Embodied Model GO-3 Preview, Plans Full Open Source by End of 2026

AgiBot has publicly released its new-generation robot embodied model, GO-3 Preview. This model connects the entire pipeline from early-stage training and simulation practice to continuous learning after deployment on the job. It has already taken first place in three authoritative industry benchmarks—RoboDojo, RoboTwin, and RoboColiseum—and is planned to be fully open-sourced by the end of 2026. For a robot to do work in the real world, being able to "think" alone is far from enough. Take small tasks like grasping a cup or assembling building blocks: cups are round or square and placed in different positions, and the robot must understand the environment in front of it, anticipate whether it will knock things over, and then perform the action steadily—this is the challenge that embodied intelligence must solve. GO-3 Preview's approach differs from the past: it is not trained once and for all in a laboratory, but instead learns gradually and keeps improving in three stages, much like a human. The first stage builds a foundation and learns to generalize. It learns extensively from videos of everyday human operations as well as records of real robot operations, and understands the relationship between objects and actions. Without separate tuning, it can attempt to complete tasks in unfamiliar scenarios such as kitchens and homes; given a small number of demonstrations, it can even take on complex, fine-grained tasks with many steps, such as making pancakes or assembling Lego. After building that foundation, it enters the second stage: repeated practice and trial and error. AgiBot has built a hybrid training system in which hundreds of robots share successful experiences and learn from failures with one another, while also pairing with simulation environments and virtual world models to simulate a wide variety of scenarios at low cost. Just like playing tabletop curling, after only 4 hours of practice the robot's skill level can surpass that of an ordinary person; when pulling blocks from a tower, it can also steadily pull out multiple blocks without letting the tower fall. The most striking breakthrough is that after the robot officially starts work, it can still evolve on its own. In the past, once many robots encountered new situations they had never seen, engineers had to come on site to re-tune parameters and collect data, making large-scale deployment difficult. Relying on this new framework, the robot automatically records successful and failed experiences while working, analyzes on its own where problems occurred, and continuously adjusts its actions without a human constantly following up with tuning. In a real convenience-store scenario test, the Genie G2 robot achieved this kind of autonomous learning: the task success rate was only 28.6% on the first day, and rose steadily to 85.7% five days later. For example, in the task of hanging small hooks, it continuously explored and adjusted the order of actions, achieving a success rate 17.7 percentage points higher than that of traditional training methods, and it can keep learning continuously without easily hitting a capability ceiling. This is also one of the rare cases in the industry where this autonomous evolution approach has been validated outside the laboratory in a real store scenario. From the early technical exploration of the GO series models to this GO-3 Preview achieving full-pipeline learning, AgiBot hopes to break the limitation of robots being "fixed once trained." The results on the three major leaderboards confirm its general manipulation capability. When robots can continuously accumulate experience and optimize themselves in real work, in the future they will no longer be merely demonstration exhibits at trade shows; they will truly enter stores and homes, continuously unlock new skills, and move step by step toward large-scale application.