GO-3 Preview: Embodied AI Enters the Era of Full-Chain Scaling

Editor | Clio

If we consider embodied intelligence as a cultivation game, the 2026 version update focuses on the 'growth system'.

This year, the world model continues to evolve, enabling robots to rehearse actions and explore the future in virtual environments; RSI (recursive self-improvement) research is also gaining momentum, attempting to convert the successes and failures of interactions into learning experiences.

From learning skills by watching videos, to repeatedly practicing in virtual and real environments, and then continuing to accumulate experience in real work scenarios, robots' 'leveling up' methods are becoming increasingly diverse, and a growth path spanning pre-training, post-training, and after deployment is becoming increasingly clear.

On October 29, AGIBOT officially released

Pre-training, fine-tuning, and post-deployment achieve end-to-end large-scale learning

and secured first place in three authoritative evaluations: RoboDojo, RoboTwin, and RoboColiseum.

From GO-1's ViLLA architecture, GO-2's 'action reasoning chain,' to GO-3 Preview's end-to-end scalable learning, AGIBOT is extending embodied models' learning to a more complete lifecycle of robots.

This time, AGIBOT hopes to establish a complete growth process for robots, from learning skills, repeated practice, to working on the job. Robots are no longer just entering work after training is complete, but can continue to accumulate experience and improve their capabilities while working.

Along this route, how exactly do robots 'upgrade' step by step?

How can robots watch videos to gain skill points?

Watching videos has become an important way for robots to learn human skills. Approximately 90% of the pre-training data for GO-3 Preview comes from first-person human data, covering a broader range of scenarios and skills than robot demonstrations.

However, understanding how humans operate does not necessarily mean that robots can follow suit. Human body posture and robot joint control, end-effector trajectories cannot be directly mapped one-to-one. Robots also need to learn how to convert observed operational experience into suitable actions for themselves.

To this end, GO-3 Preview adopts a embodied native architecture

WRAM（World Reasoning Action Model）

, linking task understanding, future state prediction, and action generation together, allowing robots to both predict how physical states will change and learn the actions required to achieve these changes.

Take LEGO assembly as an example. The robot must not only recognize the spatial relationship between parts and slots, determine the target position and orientation of the parts, but also control the robotic arm to perform grasping, alignment, and insertion, handling continuously changing contact relationships.

WRAM delegates this process to multiple collaborating modules.

First is

CoAE, the visual representation module for control.

It is responsible for extracting operational information such as object positions, shapes, and mutual relationships from the camera feed, while also ensuring the richness of visual information and real-time inference efficiency.

With the current scene representation, the next step is to predict the future.

SVP (Single-Step Visual Planner)

uses current observations and task conditions to predict multiple future visual states through a single forward calculation, then the inverse dynamics model combines the robot's own state to convert the prediction into executable actions.

Thus, human videos and robot trajectories can play their respective roles: the former helps the model learn how physical states change, while the latter provides training information from state changes to action control.

However, predicting the future still has an issue: for the same task, there may be multiple reasonable ways to complete it.

For example, placing a cup in a designated location, the robot can approach from the left or the right. Different operational paths can achieve the same goal, so the future states predicted by the model must be coordinated with actual action learning.

GO-3 Preview introduces

KASO (Knowledge Alignment Selective Optimization)

precisely for this purpose. It generates multiple candidate futures, evaluates the compatibility between predicted results and real actions, selects more suitable results for joint training, keeping future state prediction and action learning in coordination.

This indicates that

under given test conditions, expanding the scale of training data can improve the model's performance in operating new scenarios and new objects, and demonstrate performance gains across different robot bodies.

Specifically for GO-3, AGIBOT divides the pre-trained task capabilities into two settings: Zero-shot and Few-shot. The former tests the model's transfer ability without task-specific real machine supervision fine-tuning; the latter further improves the accuracy and coordination of complex operations through a small number of demonstrations.

According to AGIBOT, GO-3 Preview already has the capability to perform hundreds of tasks out of the box.

In the LEGO assembly demonstration, the GO-3 Preview, after a small amount of demonstration fine-tuning, can complete fine actions such as part grasping, alignment, and insertion, and continuously complete the assembly under Agent coordination.

In addition, GO-3 Preview also demonstrated cross-body operation capabilities. On the bipedal humanoid robot X2, the model can perform actions such as holding objects with both hands to open a door and kneeling to pick up objects.

Pre-training Scaling is expanding the skill range of robots, but to complete more complex physical tasks, relying solely on ``watching videos'' is not enough. Robots also need to practice repeatedly through interaction.

Interact between virtual and real, developing a sense of the physical world

If robots had their own version of the book 'Five Years of College Entrance Exams and Three Years of Simulation', the desktop curling game could be considered one of the physics problems in it.

The question seems simple: control the force and angle to push the opponent's curling stone out of the scoring circle, while keeping your own curling stone inside. But in practice, friction, collision, and motion trajectory all affect the outcome. A slight deviation in the force applied can change the final position of the curling stone.

According to the experimental case provided by AGIBOT, GO-3 Preview achieved a higher win rate than ordinary human players in the desktop curling task by continuously practicing in real environments and world models, within just a few hours.

This corresponds to an important direction in post-training for embodied intelligence:

Let robots continuously improve their ability to execute complex actions through interaction, feedback, and strategy updates.

However, robot practice consumes real resources. Each physical interaction involves device operation, environment reset, and data collection. Repeated experiments also involve equipment wear and safety costs. To enable robots to master more skills, it is necessary to expand the scale of interaction experience and improve data utilization efficiency.

To this end, GO-3 Preview built

Genie Evolver, a distributed hybrid environment post-training system that combines real devices, physical simulation, and world models.

The three types of environments each have their own roles.

The first type is real robots, providing reliable physical feedback.

Genie Evolver relies on a cluster of 100-level real devices, using CEL (Collective Experience Learning) to aggregate execution experience.

For example, when different robots grasp objects, some succeed while others fail. The system retrieves successful action segments, analyzes failed contexts, and updates shared strategies, allowing individual experience to serve group learning.

The benefits of this mechanism are also reflected in real device scale experiments. Under the same training time, as the scale of participating real devices increases, the success rate of the phone packaging task rises from 50.8% to 98.3%. In multi-stage phone packaging long process testing, the number of failures in 120 tests drops from 25 to 2, a reduction of 92%.

The second type is physical simulation, providing a large number of repeatable practice opportunities.

In the simulation environment, the system can adjust parameters such as object positions, friction coefficients, and collision conditions, allowing robots to try different strategies. For desktop curling, simulation can cover more initial states and throwing conditions, making it easier to compare the results of different strategies.

The third type is the world model, further expanding the space for exploring candidate actions.

Still taking curling as an example. If a throw deviates from the target, the world model can return to a previous key decision point, predicting the results of different forces and angles. In this way, a single real experience can become the starting point for exploring multiple candidate actions.

The system uses the feedback generated by the world model to optimize strategies, then verifies them in the real environment, thereby improving the efficiency of real interaction data utilization. Since the world model may still misjudge physical processes such as friction and collision, real device verification is indispensable.

Desktop curling tests precise control, while another experiment extends the difficulty to a continuously changing physical environment.

Taking the task of pulling out building blocks as an example, GO-3 Preview needs to sequentially pull out the stacked red blocks. However, each operation changes the position, posture, and force relationships of the remaining blocks. The robot must continuously observe and adjust subsequent actions, testing its ability to track and adapt continuously in long-term tasks.

From a system design perspective, the goal of Genie Evolver is to expand the scale of effective interaction that robots can utilize:

Real devices provide physical feedback, simulation expands the scope of training, and the world model explores candidate futures, with all three jointly driving continuous improvement in strategies.

And when robots complete their practice and enter more complex real work environments, these tasks themselves may also become new learning opportunities.

The robot goes to a convenience store, working while ``gaining experience''

In the game, after completing the training level, the character will often enter a new map, accept missions, and accumulate experience points.

For the GO-3 Preview, convenience stores are such a 'new map'.

Robots need to identify products, plan the order of operations, and complete picking, moving, and placing. However, shelf space, product positions, and packaging materials vary, and sliding or changes in placement space may also occur.

How to use these operational data for the next round of learning is the focus of Scaling after the deployment of GO-3 Preview.

To this end, AGIBOT built an embodied Agent framework

Genie Master, organizing task planning, execution checking, anomaly recovery, and experience memory together.

To ensure that task judgment is truly translated into action, Genie Master introduced the embodied execution protocol EEP, connecting capability invocation, execution guidance, result feedback, and task continuity. After the robot executes, information such as whether it was successful, where it got stuck, etc., will be returned to the Agent; when a task is interrupted, the system can retain progress and re-plan from the breakpoint.

Taking convenience store restocking as an example, Genie Master can arrange operational steps, check execution results, and adjust the plan after an anomaly occurs. The accumulated success and failure experiences during operations can also participate in capability improvement through two paths.

The first is to let experience enter prompts and memory, improving task planning.

Genie Master can consolidate successful, failed, and anomaly recovery cases as memory, extracting operation sequences, key constraints, and precautions. When a robot encounters a similar task again, it can retrieve the relevant experience to assist in planning and anomaly handling.

For example, when hanging goods in a convenience store, the robot once failed to locate the hook due to the head camera not seeing clearly and label boards blocking the view. After the Agent reviewed the failure record, it adjusted the execution method: letting the left hand approach the target, using the wrist camera to supplement the close-up view, and then having the right hand complete the hanging. This handling method can also be consolidated as experience for subsequent similar tasks to call upon.

The second is to let execution and recovery data enter training, updating action strategies.

The successful and failed trajectories accumulated in real operations can also be used for action model training.

For example, when packing takeout, water bottles and grippers are easily entangled by white handles. The Agent guides the gripper to first retract and then try to reinsert into the packaging bag with more specific action prompts. The related execution and recovery data, after processing, can be used for subsequent action strategy training.

This is the first time in the industry to step out of simulation environments and laboratories, validating Physical RSI (embodied recursive self-evolution) in real business scenarios.

It allows the success, failure, and anomaly recovery experiences from real operations to become material for the next round of capability updates.

According to AGIBOT, the Sprite G2 equipped with Genie Master has already been deployed and verified in real convenience stores. Without relying on human demonstration and debugging, it continuously improves in tasks such as ready-to-eat food packaging, picking, and restocking by identifying problems, adjusting execution methods, and accumulating experience.

By this point, GO-3 Preview establishes general operational capabilities, Genie Evolver optimizes strategies through virtual and real interaction, and Genie Master allows robots to continue learning from actual work.

Conclusion

In games, character leveling usually has clear experience points and skill trees. However, real-world tasks are ever-changing, making it difficult to measure a robot's growth with a single progress bar.

The real world is like a continuously updated game, with new levels appearing every day. As robots accumulate more experience, how to effectively screen, reuse experience, and make new and old skills compatible will become an important topic for large-scale deployment.

The growth of embodied intelligence will also be reflected in the continuous accumulation and transformation of these experiences: operations become more proficient, anomaly handling becomes more confident, and the robot can adapt to unfamiliar tasks faster.

Perhaps, the true 'max level' of a robot is to make those once amazing capabilities gradually become reliable, commonplace performances in everyday life.

© THE END

For republication, please contact this official account for authorization

For submissions or media inquiries: [email protected]