What is GO-2
GO-2 (Genie Operator-2) is the second-generation embodied intelligence foundation model launched by AgiBot. It connects the "understanding-planning-execution" pipeline and bridges the gap between semantics and action. The model pioneers an "Action Chain-of-Thought" that performs reasoning and planning directly in the action space, and combines it with an "Asynchronous Dual-System" architecture to achieve stable coordination between low-frequency planning and high-frequency execution. It sets new SOTA results on multiple benchmarks such as LIBERO, and through the Genie Studio platform enables continuous learning and large-scale deployment in real-world scenarios, taking robots from "understanding" to truly "stable execution".
GO-2's Main Features
- Action Chain-of-Thought Reasoning: Performs planning and reasoning directly in the action space, generating structured high-level action sequences and decomposing complex tasks into ordered steps, achieving "think it through before acting".
- Asynchronous Dual-System Execution: The slow system (low frequency) generates "intent stream" plans, while the fast system (high frequency) tracks execution in real time and adjusts dynamically, ensuring that plans are reliably fulfilled.
- Unified Semantic-Action Modeling: Connects the vision-language-action pipeline, bridges the gap between "understanding" and "execution", and enables abstract instructions to be accurately translated into actions in the physical world.
- Closed-Loop Continuous Evolution: Leveraging the Genie Studio platform and a "pre-training + post-training + data closed loop" architecture, it continuously collects data during real interactions and optimizes the model online.
- Cross-Domain Zero-Shot Transfer: Possesses strong generalization ability from simulation to real environments, adapting to new scenarios, new objects, and new textures without retraining.
- Complex Task Execution: Supports long-horizon task planning and precise physical interaction (such as grasping, placing, and manipulating objects), maintaining stable and reliable action execution in real environments.
GO-2's Technical Principles
- Action Chain-of-Thought: Traditional embodied models usually generate low-level control signals directly from vision-language input, compressing "understanding" and "execution" into the same moment, which causes a disconnect between planning and execution. GO-2 introduces the Action Chain-of-Thought to explicitly simulate the human cognitive process: before executing an action, it first forms a clear action plan internally.
- Asynchronous Dual-System Architecture
- Slow System (Semantic Planner): Runs at a lower frequency and is responsible for continuously generating and refining high-level action plans. It does not output a one-time result, but instead provides continuous constraints in the form of an "intent stream", progressively refining from macro actions to sub-actions and forming a hierarchical action representation.
- Fast System (Action Refiner): Runs at a higher frequency, continuously receives planning guidance from the slow system, and combines real-time visual observations to generate specific control signals. The execution process is not a simple reproduction of the plan, but rather real-time alignment and local correction around the plan—automatically adjusting the arm's downward reach when the table height differs from expectations, and adjusting grip force in real time when object friction characteristics change.
- Noise-Forced Teaching Training Mechanism: To ensure the fast system can maintain stable execution under "close to correct but not perfect" planning conditions, GO-2 introduces a noisy forced teaching mechanism during the training phase. When training the execution module, the real high-level action sequence is used as a condition, while a certain perturbation is added to simulate planning errors, enabling the model to continuously follow the plan in dynamic environments.
- Overall Technical Loop: The above components form a complete pipeline: the VLM and visual encoder process multimodal input → the action chain-of-thought generates executable plans in the action space → the asynchronous dual system ensures the plan is stably fulfilled → execution data flows back to the cloud for post-training, enabling continuous evolution of the model in real environments.
How to Use GO-2
- Platform Access and Deployment: Developers need to visit the Genie Studio official websitehttps://genie.agibot.com/geniestudio接入GO-2基座模型，无需本地配置复杂的训练环境即可直接调用模型能力。
- Multimodal Task Input: Input natural language instructions and current visual observation data into the system, and the VLM module will automatically parse the task intent and understand the scene information.
- Action Chain-of-Thought Planning: Based on the action chain-of-thought, the model performs reasoning and planning within the action space, generating a multi-level structured action sequence from macro to micro as the execution blueprint.
- Asynchronous Dual-System Execution: The asynchronous dual system starts automatically. The slow system continuously provides intent flow guidance at low frequency, while the fast system tracks in real time at high frequency and dynamically adjusts execution details based on visual feedback.
- Closed-Loop Continuous Evolution: During task execution, the system automatically collects real interaction data and continuously optimizes model parameters through a cloud-based post-training mechanism, achieving closed-loop evolution from simulation to real-world scenarios.
Key Information and Usage Requirements of GO-2
- Full Name：Genie Operator-2（GO-2）
- Publisher: AgiBot
- Positioning: A new-generation embodied intelligence foundation model
- Core Technologies: Action chain-of-thought (reasoning and planning in the action space) + asynchronous dual system (low-frequency planning + high-frequency execution)
- Pain Points Addressed: Bridging the "semantic-motor gap", enabling robots to go from "understanding clearly" to "executing stably"
- Performance Metrics: LIBERO 98.5%, LIBERO-Plus 86.6%, GenieSim real environment 82.9%, comprehensively SOTA
- Academic Recognition: Accepted to both CVPR 2026 and ACL 2026
- Deployment Form: Built-in Genie Studio development platform, supporting cloud-based post-training and data closed loop
Core Advantages of GO-2
- Bridging the Gap and Achieving the Unity of Knowledge and Action: Through a unified architecture, GO-2 connects the pipeline of logical reasoning and precise action execution, thoroughly bridging the semantic-motor gap between traditional robots' "high-level understanding" and "low-level execution", achieving true "unity of knowledge and action".
- Pioneering Action Chain-of-Thought for Explicit Reasoning and Planning: GO-2 pioneers the action chain-of-thought mechanism, explicitly completing reasoning and planning within the action space and decomposing complex tasks into structured action sequences, enabling robots to shift from "doing while seeing" to "thinking clearly before acting", significantly reducing execution deviation.
- Asynchronous Dual System Ensures Stable Execution: GO-2 adopts an asynchronous dual-system architecture. The slow system continuously generates "intent flow" plans at low frequency, while the fast system follows in real time at high frequency and dynamically adjusts based on visual feedback, ensuring that high-level plans are always stably executed amid real disturbances.
- Comprehensive SOTA刷新，行业领先性能: GO-2 comprehensively刷新SOTA on multiple benchmarks including LIBERO (98.5%), LIBERO-Plus (86.6%), and GenieSim real-world environment (82.9%), significantly surpassing mainstream models such as GR00T and π0.5.
- 数据闭环进化，零样本强泛化: GO-2 supports continuous collection of interaction data in real-world scenarios for online optimization, and possesses strong zero-shot cross-domain transfer capability, achieving an 82.9% success rate in real environments while being trained solely on simulation data.
GO-2的同类竞品对比
| 对比维度 | GO-2（智元机器人） | π0.5（Physical Intelligence） | RT-2（Google DeepMind） |
|---|---|---|---|
| 核心架构 | 动作思维链 + 异步双系统 | 流匹配（Flow Matching）架构 | VLA端到端（基于PaLI-X） |
| 规划方式 | 在动作空间显式推理，生成结构化高层动作序列 | 直接生成动作，无显式中间规划层 | 直接从像素和指令映射为动作标记 |
| 执行机制 | 异步双系统：慢系统低频提供”意图流”+快系统高频实时跟随修正 | 单一生成模型端到端同时完成理解与执行 | 端到端直接输出控制信号 |
| 关键优势 | 弥合语义-运动鸿沟，规划与执行强制对齐，真实场景稳定性强 | 互联网规模视觉-语言预训练，高频动作生成能力强 | 经典VLA先驱，架构简洁，端到端训练 |
| 主要局限 | 商业闭源，需配套智元硬件生态 | 规划与执行压缩在同一时刻，真实场景稳定性待提升 | 高层推理与底层控制断层，长程任务误差累积明显 |
| LIBERO成功率 | 98.7% | 96.9% | 未公开/显著较低 |
| GenieSim真实环境 | 82.9%（零样本迁移） | 77.5% | 通常低于60% |
| 进化能力 | 支持数据闭环持续进化（预训练+后训练） | 依赖离线静态数据 | 依赖离线静态数据 |
GO-2的应用场景
- 工业制造场景: In factory production lines, GO-2 can drive robots to complete fine operations such as component assembly, quality inspection, and tool operation. Through the data closed loop, it continuously learns to adapt to different workstation characteristics, achieving zero-shot transfer from simulation training to real production lines, and reducing debugging costs during production line switching.
- 商业服务场景: Applicable to services such as guidance, cleaning, and retail restocking in supermarkets, hotels, office buildings, etc. With its cross-scenario generalization capability, it can stably adapt to different shelf layouts, lighting conditions, and crowd environments without the need to collect data specifically for each store.
- 物流仓储场景: Supports repetitive operations such as sorting, handling, and palletizing. The asynchronous dual-system architecture ensures that the robot can maintain precise movements even during high-speed motion, avoiding grasping errors or collisions caused by visual deviations. At the same time, it optimizes its adaptability to different package shapes and weights through continuous data collection.
- 具身智能科研平台: As a foundation model, GO-2 provides universities and research institutions with an out-of-the-box embodied intelligence development platform. Through Genie Studio, researchers can quickly deploy the model for algorithm validation, data collection, and secondary development, accelerating academic research and technological iteration in the field of embodied intelligence.