Article · AI translation
Thanks to IT Home reader 風見暉一 for the tip!
IT Home reported on October 30 that Zhiyuan officially launched GO-3 Preview on October 29, an embodied model that enables scalable learning across the full chain of pre-training, post-training and post-deployment.
The model uses the embodiment-native WRAM (World Reasoning Action Model) architecture to connect task reasoning, future prediction and action generation, enabling robots to understand instructions, predict what will happen next and generate executable actions.
The model recently topped three authoritative industry benchmark rankings—RoboDojo, RoboTwin and RoboColiseum—and is set to be fully open-sourced by the end of 2026.
IT Home provides the following official introduction:
Based on the WRAM architecture, GO-3 Preview advances scaling in both data and models, allowing heterogeneous experiences to contribute to shared learning. The data dimension addresses “what to learn and whether the learning is broad enough”: first-person human videos provide rich examples of object interactions and task processes, while robot trajectories establish links among observations, actions and physical outcomes. The model dimension addresses “how to absorb and use these experiences”: by strengthening learning capabilities, it organizes scattered experiences and turns them into transferable, general-purpose abilities.
After pre-training, the model is better able to transfer between tasks and adapt to complex skills, as tested in zero-shot and few-shot tasks. In household, kitchen, multi-robot collaboration and dexterous manipulation scenarios, robots can complete unfamiliar tasks directly without additional fine-tuning.
For more complex tasks, GO-3 Preview can learn further from a small number of demonstrations to complete lengthy, precise tasks such as making jianbing guozi.
Assembling Lego requires accurately grasping pieces, locating the right positions and fitting the pieces together. With a fine-tuned model, the robot can perform these precise actions and, through agent orchestration, link multiple steps to complete the assembly.
Having general-purpose abilities does not mean a robot can immediately handle every complex task. Just as people still need repeated practice after learning basic movements, robots also need to improve their strategies through continual attempts and feedback.
During the GO-3 Preview post-training stage, Zhiyuan built Genie Evolver, a distributed post-training system with hybrid environments. Drawing on 100 physical robots and combining physical simulation with a world model, it enables robots to improve their abilities through practice.
As interaction experience accumulates and action strategies are continually updated, the robots’ abilities improve as well: in tabletop curling, after 4 hours of online iteration, a robot’s performance exceeded the average of the humans tested; in a block-pulling tower game, it could remove 5 blocks in succession while keeping the entire tower stable.
In the GO-3 Preview post-deployment stage, Zhiyuan uses its embodied agent framework Genie Master to connect decision-making, execution and feedback, enabling robots to improve autonomously after deployment and continue building their capabilities during real-world work.
While working in a real convenience store, a Genie G2 equipped with the framework identified the causes of its failures, adjusted its strategies and retained what it learned without relying on human demonstrations or debugging. It continually improved on multiple complex, precise tasks, achieving continuous self-improvement with “no human in the loop.”
The robot’s success rate for completing the full task workflow rose from 28.6% on the first day to a stable 85.7% after 5 days. For example, when hanging small hooks in the convenience store, the robot independently identified problems, retained what it learned and adjusted the timing of its actions, turning failure records into an action guide. Its success rate on that single task was 17.7 percentage points higher than with conventional fine-tuning (SFT); unlike the latter, which can easily hit a plateau, it supports continuous online learning.