AGIBOT has unveiled GO-3 Preview, an embodied foundation model designed to help robots learn across pre-training, post-training, and real-world deployment. Built on the company’s World Reasoning Action Model (WRAM) architecture, GO-3 combines task understanding, future-state prediction, and action generation to support general-purpose robotic capabilities and precise manipulation. The model works alongside AGIBOT’s Genie Evolver training system and Genie Master agent framework, creating what the company calls a full-lifecycle learning architecture. AGIBOT says GO-3 Preview ranked first on the RoboDojo, RoboTwin, and RoboColiseum benchmarks, and plans to open-source the model by the end of 2026.
AGIBOT Connects Robot Training With Real-World Learning
AGIBOT’s GO-3 Preview reflects an emerging direction in robotics: developing foundation models that do not stop learning when initial training is complete. Instead, the company wants robots to improve through simulated practice, physical interaction, and experience accumulated during commercial deployment.
The model is built around WRAM, or World Reasoning Action Model, an architecture that combines task understanding, prediction of future environmental states, and action generation. The objective is to help robots anticipate the physical consequences of their movements rather than simply translate instructions into predefined actions.
That distinction matters because physical tasks introduce challenges that conventional language models are not designed to handle. Picking up a cup, inserting a component, or assembling a structure requires spatial awareness, precise positioning, and an understanding of how objects respond to contact. A robot must also adapt when objects move or environmental conditions change.
AGIBOT says GO-3 Preview addresses these requirements through three connected mechanisms: CoAE (Control-Oriented Autoencoder), SVP (Single-step Visual Prediction), and KASO. CoAE extracts visual representations relevant to robotic control, preserving information about object geometry and spatial relationships. SVP predicts future states from current observations and task conditions, while KASO selects predictions that better correspond with the actions taken during training.
Together, these components are intended to connect perception, prediction, and movement in a unified learning architecture.
Pre-Training Builds Skills That Transfer Across Tasks
GO-3 Preview’s first learning stage focuses on developing general capabilities from diverse data sources. AGIBOT says the model learns from human first-person videos and robot trajectories, combining visual examples of object interaction with records that connect observations to physical actions and outcomes.
This approach is intended to reduce dependence on task-specific training. Instead of learning one fixed sequence for handling a particular cup or component, a robot should develop capabilities that can transfer to unfamiliar objects and environments.
AGIBOT reports zero-shot and few-shot task adaptation. Zero-shot capabilities allow the model to apply existing knowledge to new tasks without task-specific supervised fine-tuning on physical robots. Few-shot adaptation uses a limited number of demonstrations to improve performance on more demanding operations.
LEGO assembly is one example presented by the company. The task involves recognizing individual pieces, grasping them, aligning their positions, and inserting them accurately. The model can use a small number of additional demonstrations and fine-tuning to adapt its manipulation capabilities, while an embodied agent coordinates the individual actions into a complete task.
The broader objective is to make pre-training a foundation for general-purpose physical skills rather than a collection of isolated task demonstrations.
Genie Evolver Combines Physical Robots and Simulation
Pre-training alone cannot prepare robots for every physical situation. AGIBOT’s second stage uses Genie Evolver, a distributed post-training system that combines a fleet of 100 physical robots with physics simulation and world models.
Its Collective Experience Learning (CEL) mechanism aggregates execution experiences across robots. Successful trajectories provide examples of effective actions, while failed attempts help identify conditions that require different strategies. The system is designed to make useful experience from one robot available to others rather than requiring each machine to learn independently.
Physics simulation expands the range of interactions available for training and testing. Robots can repeat scenarios under controlled conditions, explore alternative movements, and evaluate different strategies without relying exclusively on physical experiments.
World models add another dimension by allowing the system to explore possible outcomes from critical decision points. It can compare alternative actions, predict their consequences, and use the resulting feedback to refine action policies before testing them in the physical environment.
AGIBOT illustrates the approach through tabletop curling, block extraction, and throwing a ball into a moving basket. These tasks require different combinations of force control, object manipulation, and prediction of changing conditions.
The company claims that a robot achieved a higher win rate than typical human players in tabletop curling after several hours of combined real-world practice and world-model learning. That result is an AGIBOT-reported demonstration, not an independently established comparison of general robotic performance.
Genie Master Extends Learning Into Deployment
The third stage, post-deployment learning, is built around Genie Master, AGIBOT’s embodied agent framework. It is designed to connect task planning, execution, feedback, and experience retention so that robots can use lessons from operational work to improve future performance.
Genie Master introduces the Embodied Execution Protocol (EEP), which standardizes interfaces across four areas: capabilities, guidance, feedback, and continuity. These interfaces allow an agent to identify available robot functions, determine how a task should be performed, assess whether the outcome meets requirements, and track unfinished work.
The framework also supports two learning pathways. The first converts operational experience into long-term memory. Successful and unsuccessful attempts can be recorded, with reusable lessons consolidated for later retrieval when similar situations arise.
The second pathway uses execution trajectories and feedback to improve action policies. Adjustments made during task execution and recovery can become training material for refining future behavior.
AGIBOT demonstrated the approach with its Genie G2 robot in a convenience store environment. The company says the robot identified causes of task failures, adjusted its strategies, and accumulated experience while performing operational tasks. AGIBOT characterizes this as Physical Recursive Self-Improvement, or Physical RSI, because the system uses operational experience to support continued learning.
The concept is ambitious, but it requires careful interpretation. A robot retaining operational memories or generating new training data does not automatically mean its underlying model can safely modify itself without human oversight. The effectiveness of this approach will depend on how experience is evaluated, how policy updates are validated, and how unsafe behavior is prevented from carrying into subsequent tasks.
Benchmark Rankings and the Open-Source Question
AGIBOT says GO-3 Preview achieved first-place rankings on RoboDojo, RoboTwin, and RoboColiseum. The company reports that it led RoboDojo in both task success rate and process score, achieved the highest overall score on RoboTwin, and ranked first overall on RoboColiseum.
These benchmarks provide different evaluation settings for robotic manipulation and task execution. However, leaderboard positions alone do not establish how reliably a model will perform across different robot hardware, unfamiliar environments, or extended commercial operations. AGIBOT has not provided detailed score tables or independent validation in the announcement.
The planned open-source release could offer a more substantial test of the technology. AGIBOT says it intends to fully open-source GO-3 Preview by the end of 2026. If the release includes model weights, code, documentation, and reproducible evaluation procedures, researchers and developers will have a better opportunity to assess its generalization capabilities and reproduce its results.
The development also builds on AGIBOT’s earlier GO-1 and GO-2 models, extending its approach from initial embodied learning and action reasoning toward a system that connects training, practice, and deployment.
The wider industry is pursuing a similar shift toward physical AI, where foundation models are expected to control robots and interact with changing environments rather than simply process digital information. AGIBOT’s approach brings together data scaling, simulation, fleet learning, and deployment feedback in one architecture.
Its next challenge is demonstrating that this learning cycle produces repeatable improvements in real industrial and service environments. Reliable manipulation, safe policy updates, hardware compatibility, and measurable productivity will matter more to customers than benchmark rankings alone. If AGIBOT can validate those capabilities and deliver the promised open-source release, GO-3 Preview could become a useful contribution to the emerging ecosystem of general-purpose embodied AI.
Market Landscape
Embodied AI is developing beyond conventional robotics software toward foundation models that connect perception, reasoning, prediction, and physical action. The competitive focus is shifting from training robots for individual tasks to building systems that can generalize across environments and improve through experience.
Five developments are shaping this market:
- Embodied foundation models: General-purpose models are being developed to transfer skills across different objects, tasks, and physical environments.
- World models: Predictive systems help robots evaluate possible future states before executing physical actions.
- Simulation-based training: Virtual environments expand training scenarios while reducing dependence on physical testing.
- Fleet learning: Shared experiences from multiple robots can improve collective capabilities and reduce duplicated training.
- Post-deployment adaptation: Operational feedback is becoming a potential source of continuous model and policy improvement.
AGIBOT’s approach is notable because it connects all five areas through GO-3 Preview, Genie Evolver, and Genie Master. Its commercial significance will depend on whether the architecture can deliver consistent performance outside controlled demonstrations.
Top Insights
- GO-3 Preview combines task understanding, future-state prediction, and action generation through AGIBOT’s WRAM embodied foundation model architecture.
- Genie Evolver combines 100 physical robots, physics simulation, and world models to improve robotic action policies through interaction.
- Genie Master extends learning into deployment through task execution, feedback, memory retention, and action-policy refinement.
- AGIBOT reports first-place results on RoboDojo, RoboTwin, and RoboColiseum, although independent benchmark validation remains important.
- The company plans to open-source GO-3 Preview by the end of 2026, potentially enabling broader evaluation and developer adoption.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
