GO-2 - A large-scale model of an embodied intelligent base launched by Zhiyuan Robotics. - AiBoss

What is GO-2?

GO-2 (Genie Operator-2) is the second-generation embodied intelligence platform model launched by Logic Robotics. It can bridge the gap between semantics and actions by connecting the "understanding-planning-execution" link. The model pioneered the "action thinking chain" to reason and plan directly in the action space, and combined with the "asynchronous dual-system" architecture to achieve stable coordination between low-frequency planning and high-frequency execution. It has set state-of-the-art (SOTA) records in multiple benchmark tests such as LIBERO, and can achieve continuous learning and large-scale deployment in real-world scenarios through the Genie Studio platform, enabling robots to truly move from "understanding" to "stable execution".

GO-2's main functions

Action-based reasoningIt performs planning and reasoning directly in the action space, generates structured high-level action sequences, breaks down complex tasks into ordered steps, and realizes the goal of "thinking things through before acting".
Asynchronous dual-system executionThe slow system (low frequency) generates the "intent flow" plan, while the fast system (high frequency) tracks and executes it in real time and adjusts it dynamically to ensure that the plan is stably realized.
Semantic-Action Unified ModelingIt bridges the gap between "understanding" and "execution" by connecting the vision-language-action chain, enabling abstract instructions to be accurately translated into physical world actions.
Closed-loop continuous evolutionBased on the Genie Studio platform, through a "pre-training + post-training + data closed loop" architecture, data is continuously collected and the model is optimized online in real-world interactions.
Cross-domain zero-shot transferIt has strong generalization ability from simulation to real environment, and does not require retraining to adapt to new scenes, new objects and new textures.
Complex task executionIt supports long-term task planning and precise physical interaction (such as grasping, placing, and manipulating objects), and maintains stable and reliable action execution in real-world environments.

GO-2's technical principles

Action Chain-of-ThoughtTraditional embodied models typically generate underlying control signals directly from visual-language input, compressing "understanding" and "execution" into a single moment, leading to a disconnect between planning and execution. GO-2 introduces a motor thought chain, explicitly simulating the human cognitive process: before performing an action, a clear action plan is formed internally.
Asynchronous dual-system architecture
- Slow Systems (Semantic Planner)It operates at a lower frequency and is responsible for continuously generating and refining high-level action planning. It does not output a one-time result, but provides continuous constraints in the form of an "intent flow," gradually refining macro-level actions to sub-actions, forming a hierarchical action representation.
- Quick System (Action Refiner)It operates at a high frequency, continuously receiving planning guidance from the slow system and generating specific control signals based on real-time visual observations. The execution process is not a simple replication of the plan, but rather involves real-time alignment and local corrections around the plan—automatically adjusting the arm's downward reach when the table height is inconsistent with expectations, and adjusting the gripping force in real time when the object's frictional properties change.
Noise-driven forced teaching and training mechanismTo ensure that the fast system can maintain stable execution under planning conditions that are "nearly correct but not perfect," GO-2 introduces a noisy forced teaching mechanism during the training phase. When training the execution module, real high-level action sequences are used as conditions, and a certain perturbation is added to simulate planning errors, enabling the model to continuously follow the planning in dynamic environments.
Overall technology closed loopThe above components form a complete chain: VLM and visual encoder process multimodal input → Action thought chain generates executable plans in action space → Asynchronous dual system ensures that the plans are stably implemented → Execution data flows back to the cloud for post-training, enabling the model to continuously evolve in the real environment.

How to use GO-2

Platform access and deploymentDevelopers need to access the GO-2 pedestal model through the Genie Studio website https://genie.agibot.com/geniestudio. They can directly call the model's capabilities without needing to configure a complex training environment locally.
Multimodal task inputWhen needed, input natural language commands and current visual observation data into the system, and the VLM module will automatically parse the task intent and understand the scene information.
Action Thinking Chain PlanningThe model performs reasoning and planning within the action space based on the action thought chain, generating a multi-level structured action sequence from macro to micro as an execution blueprint.
Asynchronous dual-system executionThe asynchronous dual system starts automatically. The slow system provides continuous intent flow guidance at a low frequency, while the fast system tracks in real time at a high frequency and dynamically adjusts execution details in conjunction with visual feedback.
Closed-loop continuous evolutionDuring task execution, the system automatically collects real interaction data and continuously optimizes model parameters through a cloud-based post-training mechanism, achieving a closed-loop evolution from simulation to real-world scenarios.

Key information and usage requirements for GO-2

Full nameGenie Operator-2 (GO-2)
PublisherZhiyuan Robot
positionLarge-scale model of a new generation of embodied intelligent base
Core technologiesAction-based thought chain (reasoning and planning in action space) + asynchronous dual system (low-frequency planning + high-frequency execution)
Solving pain pointsBridging the semantic-motor gap, enabling robots to move from "thinking clearly" to "acting stably".
Performance indicatorsLIBERO 98.5%, LIBERO-Plus 86.6%, GenieSim real-world environment 82.9%, fully state-of-the-art (SOTA) performance.
Academic RecognitionCVPR 2026, ACL 2026 dual-top conference reception
Deployment FormBuilt-in Genie Studio development platform, supporting cloud-based training and data closure.

GO-2's core advantages

Bridging the gap and achieving unity of knowledge and actionGO-2 uses a unified architecture to bridge the gap between logical reasoning and precise action execution, completely eliminating the semantic-motor gap between the "high-level understanding" and "low-level execution" of traditional robots, and achieving true "unity of knowledge and action".
Pioneering action-based thinking chain, explicit reasoning and planningGO-2 pioneered the action thinking chain mechanism, which explicitly completes reasoning and planning within the action space, breaking down complex tasks into structured action sequences. This transforms the robot from "doing while watching" to "thinking before doing," significantly reducing execution deviations.
Asynchronous dual system ensures stable execution.GO-2 employs an asynchronous dual-system architecture. The slow system continuously generates "intent flow" planning at a low frequency, while the fast system follows at a high frequency in real time and dynamically adjusts in conjunction with visual feedback, ensuring that high-level planning is always stably executed under real disturbances.
Completely refreshed SOTA, industry-leading performanceGO-2 has achieved state-of-the-art (SOTA) performance across multiple benchmark tests, including LIBERO (98.5%), LIBERO-Plus (86.6%), and GenieSim real-world environment (82.9%), significantly outperforming mainstream models such as GR00T and π0.5.
Data closed-loop evolution, zero-shot strong generalizationGO-2 supports continuous collection of interactive data in real-world scenarios for online optimization and has strong zero-shot cross-domain transfer capabilities, achieving an 82.9% success rate in real-world environments when trained using only simulation data.

Comparison of GO-2 with similar competing products

Comparison Dimensions	GO-2 (Intelligent Robot)	π0.5 (Physical Intelligence)	RT-2 (Google DeepMind)
Core Architecture	Action-based thinking chain + asynchronous dual system	Flow Matching Architecture	VLA end-to-end (based on PaLI-X)
Planning methods	Explicit reasoning in the action space generates structured high-level action sequences.	Directly generate actions, without an explicit intermediate planning layer.	Directly map pixels and instructions to action tags
Implementation mechanism	Asynchronous dual system: The slow system provides "intent stream" at low frequency + the fast system follows and corrects in real time at high frequency.	A single generative model simultaneously performs understanding and execution end-to-end.	End-to-end direct output control signal
Key advantages	Bridging the semantic-motor gap, forcibly aligning planning and execution, and ensuring strong stability in real-world scenarios.	Internet-scale visual-language pre-training with strong high-frequency action generation capabilities.	A pioneer of classic VLA, with a simple architecture and end-to-end training.
Main limitations	Commercial closed-source requires the support of the Zhiyuan hardware ecosystem.	Planning and execution must be compressed into a single moment; stability in real-world scenarios needs improvement.	A disconnect exists between high-level inference and low-level control, leading to significant error accumulation in long-range tasks.
LIBERO success rate	98.7%	96.9%	Not disclosed/significantly low
GenieSim Realistic Environment	82.9% (zero sample transfer)	77.5%	Usually below 60%
Evolutionary ability	Supports continuous evolution through a closed data loop (pre-training + post-training).	Depends on offline static data	Depends on offline static data

Application scenarios of GO-2

Industrial manufacturing scenariosIn factory production lines, GO-2 can drive robots to complete delicate tasks such as parts assembly, quality inspection, and tool operation. Through data closed-loop continuous learning, it can adapt to the characteristics of different workstations, achieve zero-sample transfer from simulation training to the real production line, and reduce the debugging cost when switching production lines.
Commercial service scenariosIt is applicable to services such as guidance, cleaning, and retail replenishment in supermarkets, hotels, office buildings, etc. With its cross-scenario generalization capability, it can stably adapt to different shelf layouts, lighting conditions, and traffic environments without having to collect data for each store again.
Logistics warehousing scenariosIt supports repetitive operations such as sorting, handling, and palletizing. The asynchronous dual-system architecture ensures that the robot can maintain precise movements even at high speeds, 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.
Embossed Intelligent Research PlatformAs a foundational model, GO-2 provides universities and research institutions with an out-of-the-box platform for developing embodied intelligence. Researchers can quickly deploy models through Genie Studio for algorithm verification, data collection, and secondary development, accelerating academic research and technological iteration in the field of embodied intelligence.