Zhiyuan Releases Embodied Model GO-3 Preview: Can Make Jianbing Guozi, Work at Convenience Stores, Tops Three Major Lists

Thank you for the tip from IT Home netizen Fujii Hikaru .

IT Home On October 30th, it was reported that on October 29th, AGIBOT officially launched a pre-trained / post-trained / deployed model that realizes end-to-end scalable learning. Embodied models that achieve end-to-end scalable learning —— GO-3 Preview.

The model has recently topped three major industry authoritative evaluation benchmarks: RoboDojo, RoboTwin, and RoboColiseum.

This model has recently topped the three authoritative industry evaluation charts: RoboDojo, RoboTwin, and RoboColiseum. It will be fully open-sourced by the end of 2026.

IT Home attaches the official introduction as follows:

Based on the WRAM architecture, GO-3 Preview advances Scaling from both data and model dimensions, enabling heterogeneous experiences to participate in joint learning. In the data dimension, it addresses 'what to learn and whether the learning is broad enough': first-person videos provide rich object interactions and task processes, while robot trajectories establish connections between observations, actions, and physical outcomes. In the model dimension, it addresses 'how to incorporate and apply these experiences': through enhanced learning capabilities, it organizes dispersed experiences into transferable general capabilities.

After pre-training, the model exhibits stronger task transfer and complex skill adaptability, and is tested in Zero-Shot (zero-shot) and Few-Shot (few-shot) tasks. In home, kitchen, multi-robot collaboration, and dexterous manipulation scenarios, robots can directly complete unfamiliar tasks without additional fine-tuning.

Facing more complex tasks, GO-3 Preview can leverage a small number of demonstrations for further learning, completing long-range fine tasks such as making a jianbing (Chinese pancake).

LEGO assembly requires accurate part picking, precise positioning, and completion of insertion. With a fine-tuned model, the robot can perform these fine motions, and under the coordination of an intelligent agent, connect multiple steps to complete the assembly.

Having general capabilities does not mean that robots can directly handle all complex tasks. Just like humans need to practice repeatedly after learning basic movements, robots also need to improve their strategies through continuous trial and feedback.

In the post-training phase of GO-3 Preview, AGIBOT builds a distributed hybrid environment training system Genie Evolver, relying on 100 real robots, and combining physical simulation with world models, allowing robots to enhance their capabilities through practice.

As interaction experience accumulates and action strategies continuously update, the robot's capabilities also improve: in table ice hockey, after 4 hours of online iteration, its performance exceeds the average level of tested humans; in block tower pulling, it can continuously pull 5 blocks while maintaining the stability of the entire tower.

In the deployment phase of GO-3 Preview, AGIBOT uses the embodied Agent framework Genie Master to connect judgment, execution, and feedback, enabling robots to achieve autonomous improvement after deployment and continuously enhance their capabilities in real-world operations.

The Sprit G2 equipped with this framework, during its real-world convenience store deployment, did not rely on human demonstration and tuning. It was able to self-identify failure causes, adjust strategies, and accumulate experience, achieving continuous improvement across multiple complex and precise tasks, realizing continuous self-evolution without human involvement.

The robot's complete task success rate increased from 28.6% on day one to a stable 85.7% after five days. For example, a convenience store robot hanging small hooks autonomously identifies problems, accumulates experience, adjusts the timing of actions, and turns failure records into action guidelines. Its single-task success rate improved by 17.7 percentage points compared to conventional fine-tuning (SFT), differing from the latter which often encounters bottlenecks; it supports continuous online learning.