Zhiyuan Open Sources 'Trump Card' Model, Sparking Excitement in the Robotics Community

On September 23, Zhiyuan Robot announced that its general-purpose embodied foundation model GO-1 (Genie Operator-1) has been officially open-sourced on GitHub. This is another 'technical bomb' thrown by Zhiyuan into the industry following the open-source release of the AgiBot World embodied AI million-machine dataset in January this year. In just eight months, Zhiyuan has almost publicly disclosed its entire 'family assets', from data to models.

Moreover, GO-1 is the world's first general-purpose embodied AI model adopting the Vision-Language-Latent-Action (ViLLA) architecture. Its open-source release lowers the technical threshold for embodied AI and is expected to accelerate the industrial implementation of robots.

The core of GO-1 is the Vision-Language-Latent-Action (ViLLA) architecture, which bridges the semantic gap between image-text inputs and robot execution actions by introducing implicit action tokens.

In one sentence: It first abstracts 'what it sees (vision)' and 'what it intends to do (language instructions)' into a string of low-dimensional, semantically meaningful codes (Latent Action), and then translates this 'code' into specific joint movement commands for the robot (such as how joints rotate or fingers move) via a dedicated action decoder.

Compared to VLA, the ViLLA architecture is like giving the robot an inner monologue, enabling it to better understand human intentions and thus execute actions precisely to complete corresponding tasks, thereby avoiding many instances of 'irrelevant responses'.

The core advantage of the ViLLA architecture stems from the sophisticated logic of its three-layer collaborative design:

• VLM Multimodal Understanding Layer: Built on InternVL-2B, it can process multimodal information such as multi-view visual images, force signals, and natural language instructions, providing the system with scene perception and instruction interpretation capabilities. It serves as the foundation for the robot to "see the world and understand needs."
• Latent Planner: Acting as the "decision-making brain," it achieves high-level planning and understanding of complex tasks by predicting latent action tokens. For example, upon receiving the command to "hang clothes," it can first plan out the action logic of "grabbing the collar - locating the hook - adjusting posture."
• Action Expert: Relying on diffusion models to generate high-frequency, high-precision continuous action sequences, it ensures that robots can complete fine manipulation tasks such as unscrewing bottle caps and folding clothes, addressing the pain points of traditional models where actions are stiff and lack precision.

To lower the barrier for developers, AgiBot's Genie Studio one-stop development platform provides a full-stack solution for developers and partners. The platform deeply integrates AgiBot's open-source million-scale real-world and simulation datasets, offering proprietary and mainstream foundation models, complete simulation toolchains, high-fidelity simulation data, and enabling one-click deployment to physical machines.

From data collection to model deployment, the platform's GO-1 foundation model enables 'out-of-the-box' usage. It integrates Video Training solutions and a unified training framework, paired with a complete development toolchain, allowing developers to avoid piecing together tools.

Furthermore, it supports one-click compilation and deployment on real machines, which can help shorten the development cycle, improve development efficiency, and provide strong support for the faster implementation of Embodied AI technology in practical applications.

Although GO-1 was pre-trained only on data from the AgiBot G1 robot, it has demonstrated strong adaptability across various hardware and scenarios. In cross-platform verification tests, the model successfully adapted to different bases such as Unitree Robotics, Ark Robotics, and Franka robotic arms, confirming that GO-1 possesses good portability through validation of its kinematic characteristics and control interface compatibility.

Performance test data more intuitively demonstrates GO-1's technical capabilities:

• GenieSim Simulation Evaluation: Maintaining a leading position with an overall score of 3.793, it showed significant advantages in tasks such as packing items moving on conveyor belts, restocking supermarkets, and packing supermarket goods, demonstrating strong adaptability to complex scenarios.

• Libero Simulation Evaluation: It maintained advantages in sub-dimensions such as Libero Spatial and Libero Object, confirming its core capabilities in spatial perception and object manipulation.

• Real-Robot Experiments: In real-robot experiments using the Genie G1, GO-1 performed outstandingly in daily tasks such as refilling drinks, folding shorts, and clearing tables, maintaining a lead compared to other SOTA models.

From open-sourcing datasets to open-sourcing models, Zhiyuan Robot is building a low-barrier, highly collaborative Embodied AI ecosystem.

Notably, shortly before this, Zhiyuan officially launched the Genie Trailblazer global recruitment program, inviting Embodied AI research teams worldwide to focus on three core exploration directions: general Embodied AI models, Embodied World Models, and advanced teleoperation.

With the comprehensive open-sourcing of Zhiyuan Robot's GO-1 general Embodied foundation model, it is evident that Zhiyuan is pushing the construction of the Embodied AI ecosystem into deeper dimensions through a dual strategy of "resource openness + talent aggregation."

It is not difficult to foresee that with the deep integration of high-quality data, mature models, and top-tier global research forces, the future when Embodied AI truly enters people's lives will not be far off.