China’s embodied-AI race is increasingly focused on a less glamorous but decisive input: the volume and quality of robot data. AgiBot has released GE-Act 2.0, a world-action model for robotic manipulation that it says was trained from random initialization on embodied data. Pandaily reported on September 10 that the company scaled its training set from 300 to 30,000 hours and observed stronger zero-shot performance on tasks including towel folding, nesting paper cups, and flower arranging. The release is a research result, not proof of a commercial robot deployment, but it offers a concrete view of how data scale may affect manipulation models.
AgiBot describes GE-Act 2.0 as a “native” world-action model. The label refers to an architecture that connects observation, prediction, and action rather than treating vision and robot control as entirely separate systems. The company’s materials say the model learns from embodied data and uses a visual planner to anticipate future states before mapping them into actions. That is a significant claim because robotic manipulation requires a system to understand not only what an object looks like, but also how it may move, deform, or respond when touched.
The reported experiments also fit a broader effort to reduce dependence on heavily scripted robot behavior. EastFrontier recently covered Qianjue’s funding for a robot world-model push. GE-Act 2.0 is centered on manipulation rather than on video generation or a general simulation platform. It reflects a shared effort to give robots models that can predict consequences, not only recognize scenes.
From 300 to 30,000 Hours of Embodied Data
AgiBot’s central result is a 100-fold scaling experiment. Its researchers moved from 300 hours of training data to 30,000 hours and reported that certain zero-shot tasks became possible only at the larger scale. “Zero-shot” means a model attempts a task without being trained specifically on that exact example. In the reported demonstrations, the model handled activities such as folding a towel, putting paper cups inside one another, and arranging flowers. These are developer-reported results, so they should be understood as evidence from AgiBot’s own evaluation rather than third-party certification.
The distinction matters. Manipulation benchmarks often depend on robot hardware, camera position, object properties, task definitions, and how failures are counted. A model that works in a controlled laboratory may still fail when lighting changes, objects are damaged, or a human steps into the workspace. AgiBot’s release is valuable because it specifies a data-scale comparison, but it does not establish that the system is ready for unsupervised work in a warehouse, factory, or home.
The company reports that performance on zero-shot, out-of-distribution tasks increased from 17.1% to 44.1% on its G1-OP platform and from 13.4% to 31.1% on G2-90D as the data set scaled. Those numbers give the announcement more substance than a visual demo alone. They also show the distance still remaining. Even the higher reported rates leave many tasks unsuccessful, which is a reminder that larger training sets improve performance without eliminating the reliability problem.
GE-Act 2.0 Links Visual Prediction to Robot Actions
The GE-Act 2.0 project page says the system uses a CoAE visual encoder to compress 256-by-384-pixel frames into 24 tokens. It then uses a visual planner to predict future states and translate them into action sequences. The specific architecture may change with later versions, but the underlying approach is clear: make visual anticipation a central part of control rather than asking a robot to react one step at a time.
This design helps explain why the model is called a world-action system. A conventional robot-control pipeline can detect an object, estimate its position, and command a gripper. A model that predicts what will happen after an action could potentially choose sequences that account for movement and contact. The payoff is not guaranteed. Predicting future visual states accurately becomes harder as tasks become longer or involve soft, reflective, or cluttered objects. Still, it is a plausible route toward more adaptable manipulation.
AgiBot also reports approximately 104 milliseconds of latency per action chunk when the full pipeline runs on a single Nvidia RTX 5090 graphics processor. This is a reported hardware result, not a generalized latency promise. It does indicate that the team is measuring inference speed alongside task success, an essential consideration for machines that need to react in the physical world. A system with strong benchmark accuracy but excessive latency may be unusable around moving people or production equipment.
China’s hardware and data ecosystem is becoming increasingly relevant to this kind of research. The country’s robot makers need models, but they also need the recordings, sensors, processors, and test environments to train them. EastFrontier’s report on Wuwen AI’s physical-AI data financing highlighted how specialized data companies are trying to supply that resource. AgiBot’s 30,000-hour experiment provides a numerical illustration of why those data businesses matter.
More Data Improves Results but Does Not Solve Deployment
The public release should not be read as a claim that humanoid robots have reached general-purpose autonomy. AgiBot’s materials identify improved performance on selected platforms and tasks. They do not provide a customer deployment, a safety record, a cost per trained hour, or an independently reproduced benchmark. Those omissions are normal for a research announcement, but they define its limits.
The most useful takeaway is narrower. Scaling embodied data can make a measurable difference in a model’s ability to handle unfamiliar manipulation tasks. That conclusion supports the investment thesis behind China’s data-collection facilities, robot fleets, and simulation systems. It does not tell investors which company will build the most reliable robot, because reliability depends on hardware, control systems, evaluation, economics, and application fit as well as model training.
AgiBot now needs to show whether its reported gains transfer to environments that matter commercially. Warehousing, manufacturing, retail, and service settings all contain tasks where objects vary and errors have a cost. A model that can manipulate a towel in a controlled evaluation may still need more data, better sensing, and careful safeguards before it handles customer inventory or works near people.
GE-Act 2.0 therefore adds a useful piece of evidence to the physical-AI debate. Data scale is not a guarantee of general intelligence, but the results suggest that it can unlock capabilities that smaller collections did not produce. The next question is whether AgiBot and its peers can turn that learning curve into systems that are accurate, responsive, and affordable enough to leave the demonstration floor.