Predict the future, then choose the action
GE-Act 2.0 is AGIBOT's native world-action model for robot manipulation. Its central result is not that manipulation is solved; it is that zero-shot performance kept improving as embodied data grew from 300 to 30,000 hours.
Quick answer
GE-Act 2.0 connects future-state prediction with robot action generation and trains both for embodied manipulation from scratch. In AGIBOT's zero-shot evaluation, scaling the training set 100× increased both the number of tasks with any success and the overall success rate across two robot embodiments—but the best reported overall rate was still 44.1%, so this is evidence of a scaling path, not general-purpose manipulation.
What is a world-action model?
A conventional action model can be described, roughly, as a system that observes the current scene and predicts the robot action that should come next. A world-action model adds an intermediate question: if the robot acts in a certain way, what is the world likely to look like afterward?
That matters in manipulation because a useful action is defined by its effect. Moving a gripper twelve centimetres is not the goal; grasping the cup without knocking it over, predicting where it will move and placing it in a stable final pose are the goal. The model therefore needs a relationship between the visible world, a plausible future state and the motor commands that could produce that state.
Useful distinction: a world-action model is not simply a video generator connected to a robot. The future prediction must preserve the object, motion and spatial details needed to generate an executable action.
GE-Act 2.0 was trained for robotics from the beginning
AGIBOT calls GE-Act 2.0 a native world-action model because the visual representation, future prediction and action prediction are trained from random initialization on embodied manipulation data. The company contrasts this with approaches that begin from a pretrained video-generation model and add an action head later.
The claim is architectural, not mystical. Training the world model and action model together gives action errors a route back into the future-state predictor. The predicted future can be optimized around what the robot actually needs to execute, instead of only around whether a generated video looks plausible to a person.
“Native” does not mean the model understands every physical situation, nor does it make every prediction accurate. It means the representation and planning stack were developed jointly around embodied action rather than inherited as a finished visual model from another task.
What did AGIBOT test from 300 to 30,000 hours?
AGIBOT trained GE-Act 2.0 with the same reported training configuration at four data scales: 300, 1,200, 5,000 and 30,000 hours. The evaluation covered 100 atomic tasks, 20 skill categories and two robot embodiments, identified as G1-OP and G2-90D.
The evaluation was zero-shot. AGIBOT says the model was tested without task-specific fine-tuning, extra demonstrations or replacement, and that the test scenes used unseen backgrounds, lighting conditions and object instances. That design matters because it asks what the pretraining itself produced, rather than how well the model can rehearse each benchmark after seeing it.
There are two separate questions in the results. First, did success improve on tasks the model could already attempt? Second, did tasks with zero success at smaller scales begin to show non-zero success as the training set grew? The published results answer both in the affirmative, within AGIBOT's own benchmark and conditions.
More tasks became possible, and average success increased
At the two ends of the 100× scaling run, AGIBOT reported improvements on both embodiments. G1-OP moved from 39 to 76 tasks with non-zero success, while G2-90D moved from 24 to 72. Overall success rose from 17.1% to 44.1% on G1-OP and from 13.4% to 31.1% on G2-90D.
G1-OP
- Tasks with non-zero success
- 39 → 76
- Overall success rate
- 17.1% → 44.1%
- Data scale
- 300 h → 30,000 h
G2-90D
- Tasks with non-zero success
- 24 → 72
- Overall success rate
- 13.4% → 31.1%
- Data scale
- 300 h → 30,000 h
What 44.1% means: it is the manufacturer-reported overall success rate for G1-OP in this zero-shot evaluation. It does not mean GE-Act 2.0 has solved 44.1% of all manipulation, nor does it establish the reliability required for a factory deployment.
AGIBOT also says the trend did not show a clear sign of saturation between 5,000 and 30,000 hours. That is evidence that more data still helped under the tested setup. It is not proof that the same curve will continue indefinitely, and the announcement does not provide an independent replication.
What does “emergent” mean in this experiment?
AGIBOT lists folding towels, nesting paper cups, uncapping and recapping pens, and arranging flowers among the tasks that could not be completed at smaller data scales but began to show success at larger scales. In this context, “emergent” has a narrow meaning: a task went from zero observed success to non-zero observed success as the pretraining data increased.
That is useful evidence of scaling behaviour because the larger model was not given a special fine-tuning pass for each listed task. It suggests that additional embodied experience improved reusable manipulation structure rather than only memorizing the earlier task set.
It is not evidence that the robot acquired a human-like concept of folding or flower arranging. Non-zero success can still be low, brittle or limited to the evaluation setup. A deployment decision would need per-task success rate, object variation, cycle time, recovery behaviour and performance across repeated customer-operated trials.
Can experience from one robot body help another?
G1-OP accounted for more than half of the training data, while G2-90D represented less than 2%, according to AGIBOT. Despite that imbalance, G2-90D's overall success rate improved by 17.7 percentage points as the shared dataset grew.
AGIBOT interprets this as evidence that some capability learned from data-rich embodiments can transfer to an embodiment with much less direct data. If that pattern continues, it could reduce the need to collect a completely isolated dataset for every arm, hand, camera position or joint configuration.
The limitation is important: “transfer” does not mean the two bodies become interchangeable. Geometry, reach, payload, sensing, end effectors, control frequency and safety limits still change what an action means. The result supports shared learning across embodiments; it does not remove the need for body-specific validation.
CoAE, visual planning and KASO—without the paper jargon
GE-Act 2.0 combines three ideas that address different bottlenecks: representing an image efficiently, predicting the future quickly enough to act and avoiding bad training signals when several futures could all be valid.
CoAE
The control-oriented autoencoder compresses each 256×384 frame into 24 visual tokens. AGIBOT says this is one-sixteenth the token count used by DINOv3 while retaining action-relevant semantics, motion and local structure. In a controlled set of more than 4,000 manipulation samples, AGIBOT reports 97.95% instruction-image matching accuracy.
Native visual planning
A world model predicts a future state, then an inverse dynamics model generates corresponding actions. GE-Act 2.0 uses single-step visual planning, and AGIBOT reports that a chunk of continuous actions can be generated in about 104 milliseconds on an NVIDIA GeForce RTX 5090. That figure is hardware- and setup-specific.
KASO
Some tasks have several valid solutions, such as grasping with either hand. KASO generates multiple possible futures and selects the one most consistent with the demonstrated action. AGIBOT reports a 7.5-point gain in target-selection accuracy and a 10-point gain in final grasp success in its out-of-distribution evaluation.
These components do not eliminate model error. They explain how AGIBOT tries to make future prediction more compact, action-relevant and trainable. The practical question remains whether the resulting policy meets a defined task's reliability and safety threshold on the exact robot and site.
Why failed attempts and human video can still be useful
The 300-to-30,000-hour scaling experiment should not be confused with every data quantity used elsewhere in the GE-Act 2.0 training pipeline. AGIBOT separately reports 39,000 hours of instruction-video data for world-model pretraining, 32,000 hours of robot trajectories for inverse-dynamics pretraining and 30,000 hours of complete instruction-video-action data for joint training.
The world-model data includes 3,000 hours of embodiment-free material such as egocentric and human-manipulation video without robot action labels. The robot trajectories include 2,000 hours of failed manipulation attempts and deployment rollouts without success labels.
The reason this can work is modular supervision. Video without robot actions can still teach the world model how objects change. Robot trajectories without instructions can still teach the inverse dynamics model which actions correspond to visual change. Fully labelled demonstrations connect the pieces. The approach potentially turns more of a robot programme's real operating history into training material, but data governance, privacy, version control and failure review remain deployment responsibilities.
GE-Act 2.0 and AGILE 2.0 solve different layers
GE-Act 2.0 and AGILE 2.0 should not compete for the same explanation. GE-Act 2.0 addresses manipulation intelligence: predicting future states and generating object-directed actions. AGILE 2.0 addresses motion control: using perception to reshape locomotion, balance and whole-body movement while execution is underway.
| Question | GE-Act 2.0 | AGILE 2.0 |
|---|---|---|
| Primary layer | Manipulation intelligence | Locomotion and whole-body control |
| Core question | What should the robot do next to change the world? | How should the body move and recover as conditions change? |
| Key mechanism | Future-state prediction plus action generation | Perception-driven closed-loop motion |
| Representative tasks | Grasping, folding, placing and object interaction | Balancing, walking, recovering and loco-manipulation |
| Evidence in the announcement | Zero-shot scaling across 100 tasks and two embodiments | Prepared X2 demonstrations and control-architecture description |
For the locomotion side, read SpeedyDrone's separate AGIBOT AGILE 2.0 explainer. Keeping the two layers separate makes it easier to ask whether a project needs better action generation, better body control or both.
What GE-Act 2.0 does not prove
The announcement is meaningful research evidence, but it is not a finished deployment specification. A responsible reading preserves the difference between a benchmark trend, a software package and a commercial robot delivered to a customer.
- It does not establish general intelligence or human-level physical understanding.
- It does not show that every manipulation task can be completed.
- It does not make a 44.1% aggregate success rate equivalent to production reliability.
- It does not prove long-duration uptime, recovery or acceptable cycle time.
- It does not establish that every commercial G1 or G2 includes GE-Act 2.0.
- It does not map G1-OP or G2-90D directly to every current retail configuration.
- It does not remove human supervision, site safety or manual fallback requirements.
- It does not replace a customer-run acceptance test with the exact objects and environment.
Research embodiments are not automatically retail configurations
AGIBOT identifies G1-OP and G2-90D as the two embodiments in the scaling evaluation. The public announcement does not provide enough configuration detail to treat those names as identical to every current G1 or G2 SKU listed by SpeedyDrone. Buyers should not assume that GE-Act 2.0, its training stack, benchmark policies or model weights ship with a commercial platform unless the written project scope says so.
SpeedyDrone's current G Series listings do provide two relevant project directions. AGIBOT G1 is positioned around embodied-AI data collection, teleoperation, model development and inference. AGIBOT G2 is positioned around higher-precision, force-controlled manipulation and industrial operation. Those are product roles, not proof of GE-Act software inclusion.
AGIBOT G1
Use G1 as the starting point when the project centres on collecting embodied data, teleoperating the robot, developing models or evaluating inference across controlled environments. Confirm the exact end effector, data tools, software access and integration scope.
View AGIBOT G1 at SpeedyDrone Canada
AGIBOT G2
Use G2 as the starting point when the project needs more demanding physical manipulation, force-aware handling or deployment-oriented hardware. Confirm compute version, end effectors, tooling, software, safety controls and support in the quotation.
View AGIBOT G2 at SpeedyDrone CanadaWhat a Canadian robotics team should verify before procurement
A scaling result can justify a technical conversation; it cannot define the acceptance criteria. Start with one observable manipulation workflow and specify the exact environment, objects, success condition and intervention policy before selecting hardware or software.
- Which object set, grasp types, tolerances and failure modes define the task?
- Which robot embodiment, end effector, sensors and compute configuration will be delivered?
- Is GE-Act 2.0 included, licensed, accessible for development or only referenced as AGIBOT research?
- Where does inference run, and what happens if compute, network or perception fails?
- What consecutive-cycle success rate, cycle time and intervention rate are acceptable?
- How will the system recover, stop safely and hand control back to an operator?
- Who owns data, model versions, updates, cybersecurity and change approval?
- What Canadian workplace, privacy, accessibility and site-safety review is required?
For broader evidence about manipulation and scenario execution, SpeedyDrone's AGIBOT World Humanoid Robot Games analysis explains what competition results can—and cannot—tell a buyer.
Questions robotics teams are likely to ask
What is AGIBOT GE-Act 2.0?
GE-Act 2.0 is AGIBOT's native world-action model for embodied manipulation. It combines a world model that predicts a future visual state with an inverse dynamics model that generates robot actions corresponding to that change. AGIBOT trained its visual representation, future prediction and action prediction from random initialization on embodied data rather than starting from a finished video-generation model.
What is the difference between a world model and an action model?
A world model predicts how a scene may change, while an action model predicts what the robot should do. A world-action model connects the two: it predicts a plausible future state and then generates actions that could move the current scene toward that state. The connection is useful only when the visual representation preserves the object and motion details needed for execution.
Why does AGIBOT call GE-Act 2.0 “native”?
AGIBOT uses “native” to mean that the representation, world prediction and action prediction were trained together for embodied manipulation from scratch. The company contrasts this with adapting a pretrained video generator by adding an action head. The term describes the training architecture; it does not mean the model is universally capable or automatically included with every AGIBOT robot.
What does zero-shot manipulation mean here?
In AGIBOT's evaluation, zero-shot means the pretrained model was tested without task-specific fine-tuning, extra demonstrations or model replacement. AGIBOT also says the test scenes used unseen backgrounds, lighting conditions and object instances. The purpose is to measure what the shared pretraining learned, but zero-shot performance still depends on the exact benchmark, task definitions and evaluation conditions.
Did GE-Act 2.0 solve general robot manipulation?
No. The strongest reported overall zero-shot success rate in the scaling comparison was 44.1% on G1-OP at 30,000 hours. That is a meaningful improvement over 17.1% at 300 hours, but it remains far from a universal reliability claim. The results support a data-scaling trend within AGIBOT's evaluation; they do not establish production readiness across arbitrary tasks and environments.
What skills emerged as the training data increased?
AGIBOT lists folding towels, nesting paper cups, uncapping and recapping pens, and arranging flowers among the tasks that moved from zero success at smaller scales to non-zero success at larger scales. “Emergent” therefore means newly observed success in the benchmark, not spontaneous human-like learning or guaranteed mastery of those tasks.
What is cross-embodiment transfer?
Cross-embodiment transfer is the ability for learning from one robot body to help another body with different geometry or representation in the dataset. AGIBOT reports that G2-90D represented less than 2% of the training data yet improved by 17.7 percentage points as the shared dataset grew. The result suggests partial transfer, not that different robots become interchangeable.
How is GE-Act 2.0 different from AGILE 2.0?
GE-Act 2.0 focuses on manipulation intelligence: predicting future states and generating actions for object interaction. AGILE 2.0 focuses on locomotion and whole-body control: using perception to keep reshaping movement, balance and contact decisions while the robot acts. A complex embodied workflow may need both layers, but the announcements provide different evidence and should not be merged into one capability claim.
Is GE-Act 2.0 included with AGIBOT G1 or G2?
The public GE-Act 2.0 announcement does not establish that every commercial G1 or G2 includes the model, its weights, development tools or every benchmark capability. It names G1-OP and G2-90D as research embodiments, but does not map them to every retail configuration. Canadian buyers should confirm the exact robot, end effector, software package, model access, licences, integration and support in the written project scope.
Primary evidence and related SpeedyDrone guidance
AGIBOT GE-Act 2.0 announcement
AGIBOT Releases GE-Act 2.0, Providing the First Systematic Validation of a Scaling Path for Native World-Action Models, September 10, 2026. Source for architecture, datasets, evaluation design and manufacturer-reported results.
Current SpeedyDrone G Series records
AGIBOT G1, AGIBOT G2 and the AGIBOT G Series collection. Hardware, software inclusion, configuration and availability should be reconfirmed for the exact project.
Adjacent control and deployment evidence
SpeedyDrone's AGILE 2.0 explainer covers locomotion intelligence; the World Humanoid Robot Games analysis covers separate competition evidence.
Start with the manipulation task, then choose the platform
Compare G1 and G2 around data collection, teleoperation, force control, end effectors, model access and deployment requirements—then confirm the exact GE-Act 2.0 scope in writing.
Explore the AGIBOT G Series
