What is GO-1
GO-1 (Genie Operator-1, AgiBot Qiyuan Large Model) is the first general-purpose embodied foundation model launched by AgiBot. The model adopts the Vision-Language-Latent-Action (ViLLA) architecture and consists of a VLM (multimodal large model) and a MoE (mixture of experts). The VLM, leveraging massive internet image-text data, endows the model with general scene perception and language understanding capabilities; the Latent Planner in the MoE, through large amounts of cross-embodiment and human operation video data, acquires general action understanding capabilities; and the Action Expert, based on millions of real-machine data, achieves precise action execution.
Main features of GO-1
- Human video learning: By analyzing large amounts of human operation video data, the model can learn and understand action knowledge in the real world and quickly adapt to new tasks.
- Fast few-shot generalization: With very little data or zero samples, GO-1 can quickly generalize to new scenarios and tasks, lowering the threshold for embodied intelligence applications.
- One brain, multiple forms, cross-embodiment application: GO-1 can be flexibly deployed on different types of robot bodies, supports multiple robot forms, and demonstrates extremely high versatility and flexibility.
- Continuous evolution: In actual use, GO-1 can continuously learn and optimize its own performance, and through a data feedback system it continuously evolves from problem data encountered in actual execution, becoming smarter the more it is used.
- Efficient action execution: Based on the Action Expert trained on million-level real-machine data, the model has precise and efficient action execution capabilities.
Computing principles of GO-1
- VLM (multimodal large model): Through deep mining of massive internet image-text data, the VLM endows the model with excellent general scene perception and language understanding capabilities. It can accurately recognize and understand information in images while efficiently integrating with text data to achieve comprehensive understanding of complex scenarios.
- MoE (mixture of experts system): The MoE system further enhances the model's action understanding and execution capabilities. Among them:
- Latent Planner: By analyzing large amounts of cross-embodiment and human operation video data, it has mastered general action planning logic.
- Action Expert: Relying on million-level real-machine data training, it has precise and efficient action execution capabilities.
GO-1 project address
- Project official website：https://agibot-world.com/blog/go1
- GitHub repository：https://github.com/OpenDriveLab/AgiBot-World
- HuggingFace model library：https://huggingface.co/agibot-world/GO-1
- Technical paper：https://agibot-world.com/blog/agibot_go1
Application scenarios of GO-1
- Retail services: In retail environments, GO-1 can be deployed as a service robot to provide customer guidance, product inquiries, checkout assistance, and other services.
- Reception and consultation: In places such as hotels, restaurants, or office buildings, GO-1 can serve as a reception robot to provide information consultation, reservation confirmation, direction guidance, and other services.
- Production line assistance: In manufacturing, GO-1 can assist with repetitive tasks on assembly lines, such as parts handling and assembly.
- Housework assistant: In home environments, GO-1 can serve as a housework assistant to help complete daily chores such as cleaning and organizing.
- Scientific research exploration: GO-1 can be used in scientific research fields, such as sample collection and data analysis in extreme environments.