Agibot Unveils Embodied Model GO-3 Preview, Tops Three Benchmarks
Agibot has released GO-3 Preview, an embodied AI model based on the WRAM architecture that enables full-chain scalable learning. The model tops three major benchmarks (RoboDojo, RoboTwin, RoboColiseum), demonstrating zero-shot transfer and complex manipulation skills like making crepes, with a promise of full open-sourcing by late 2026.
Full Report
IT Home reported on October 30 that on October 29, AgiBot officially launched GO-3 Preview, an embodied model that achieves full-chain scalable learning across pre-training, post-training, and deployment.
The model adopts the embodied-native WRAM (World Reasoning Action Model) architecture, connecting task reasoning, future prediction, and action generation to enable robots to understand instructions, predict outcomes, and generate executable actions.
Recently, the model topped three authoritative industry benchmarks: RoboDojo, RoboTwin, and RoboColiseum. It is scheduled for full open-source release by the end of 2026.
Here is the official introduction provided by IT Home:
Based on the WRAM architecture, GO-3 Preview advances scaling from both data and model dimensions, enabling heterogeneous experiences to participate in joint learning. The data dimension addresses "what to learn and whether the breadth is sufficient": first-person human videos provide rich object interactions and task processes, while robot trajectories establish connections between observations, actions, and physical outcomes. The model dimension addresses "how to absorb and utilize these experiences": by enhancing learning capabilities, it organizes scattered experiences into transferable general competencies.
After pre-training, the model demonstrates stronger task transfer and complex skill adaptation capabilities, validated through Zero-Shot and Few-Shot tasks. In home, kitchen, multi-robot collaboration, and dexterous manipulation scenarios, robots can directly complete unfamiliar tasks without additional fine-tuning.
For more complex tasks, GO-3 Preview can leverage a small number of demonstrations for further learning, completing long-horizon fine-grained tasks such as making jianbing guozi (Chinese crepes).
LEGO assembly requires precise part grasping, accurate positioning, and successful insertion. With a fine-tuned model, robots can perform these delicate actions and, under agent orchestration, chain multiple steps together to complete the assembly.
Having general capabilities does not mean robots are immediately ready for all complex tasks. Just as humans need repeated practice after learning basic movements, robots must also improve their strategies through continuous trial and feedback.
During the post-training phase of GO-3 Preview, AgiBot built the distributed hybrid environment post-training system, Genie Evolver. Relying on hundreds of real robots and combining physics simulation with world models, this system enables robots to enhance their capabilities through practical experience.
As interaction experiences accumulate and action strategies continuously update, robot capabilities improve accordingly: in table-top curling, after 4 hours of online iteration, its performance exceeded the average level of human participants; in Jenga tower extraction, it successfully pulled out 5 blocks consecutively while maintaining the stability of the entire tower.
In the post-deployment phase of GO-3 Preview, AgiBot uses the embodied Agent framework, Genie Master, to connect judgment, execution, and feedback. This allows robots to achieve autonomous improvement after deployment, continuously enhancing their capabilities during real-world operations.
The Spirit G2 robot, equipped with this framework, worked in a real convenience store. Without relying on human demonstrations or debugging, it independently identified failure causes, adjusted strategies, and accumulated experience. It achieved continuous improvement across multiple complex fine-grained tasks, realizing "human-in-the-loop-free" continuous self-evolution.
The success rate of the robot's complete task workflow rose from 28.6% on the first day to a stable 85.7% after five days. For example, when hanging small hooks in the convenience store, the robot autonomously located issues, accumulated experience, adjusted action timing, and transformed failure records into action guides. Its single-task success rate improved by 17.7 percentage points compared to conventional Supervised Fine-Tuning (SFT). Unlike SFT, which often hits bottlenecks, this approach supports continuous online learning.
