AgiBot unveils GE-Act 2.0, testing real-robot zero-shot performance from 300 to 30,000 hours of data

Zhiyuan Innovation (Shanghai) Technology Co., Ltd. unveiled its native World Action Model (WAM), GE-Act 2.0, on Sept. 9, 2026. The company expanded training data through four stages: 300 hours, 1,200 hours, 5,000 hours and 30,000 hours. It said it conducted real-robot zero-shot evaluations across 100 atomic tasks, 20 skill categories and two robot platforms. According to the announcement, a 100-fold increase in data enabled new fine-grained actions such as folding towels, nesting paper cups, removing and replacing pen caps, and arranging flowers.

Zhiyuan Innovation showed how real-robot zero-shot success rates and the number of tasks covered changed as training data increased from 300 to 30,000 hours. This suggests robot performance evaluation is moving beyond one-off demonstrations toward learning curves that account for data volume and differences between robot platforms. According to the company, G2-90D also improved despite receiving less than 2% of the training data, making the conditions for transfer to platforms with limited data a key issue.

The core of the announcement is not the GE-Act 2.0 name itself, but the quantitative demonstration of how the range of real-robot zero-shot capabilities changes as data increases across four stages, from 300 to 30,000 hours. Rather than simply presenting performance, the company tested what scale of data is needed for different skills to emerge in a native World Action Model. At the same time, training all models from random initialization without repurposing a video-generation model extends an established approach: acquiring world prediction and action control simultaneously from embodied data. Zhiyuan subsidiary Critical Point wins seven gold medals at World Humanoid Robot Games and Zhiyuan wins eight gold, five silver and six bronze medals at robot games; official announcement limited showed Zhiyuan putting its hardware capabilities front and center through mass-produced robots in competition and their results. This time, the focus has shifted from hardware performance to explaining data scale and training methods. In other words, GE-Act 2.0 translates the capabilities demonstrated at the games into a learning curve of 300, 1,200, 5,000 and 30,000 hours, shifting the focus from demonstrations to foundation models. Structurally, the model goes beyond a simple split between perception and control. It learns changes in the world from 39,000 hours of “instruction-video” data, learns inverse dynamics from 32,000 hours of trajectory data, and connects the two using 30,000 hours of complete data. This highlights how robot capabilities depend not just on the hardware itself, but on how video, trajectories, failure data and rollouts are combined. The company’s claim that G2-90D improved with less than 2% of the data is an indication of possible transfer to platforms with limited data, while also raising the question of which platform configurations can share capabilities and to what extent. For the industry, a major implication is that competition is increasingly centered not only on “how intelligent a model is,” but also on how much diverse embodied data it can learn from and across how many robot platforms. For WAMs that integrate world prediction and action generation, data-collection design, the handling of failure trajectories and the conditions for zero-shot evaluation directly constrain commercialization. The next points to examine are the composition of the data supporting training at the 30,000-hour scale, the differences between the two robot platforms, and how the model connects to mass-produced robots and external development environments.

The company said all GE-Act 2.0 model parameters were trained from random initialization, without repurposing an existing video-generation model, and that world representations, future generation and action prediction were trained together using embodied operation data. Real-robot evaluations used zero-shot conditions in which scenes, backgrounds, lighting and object instances were excluded from training. The model was tested on two platforms, G1-OP and G2-90D. According to the announcement, the final training stage used four data scales: 300, 1,200, 5,000 and 30,000 hours. The number of tasks with nonzero success increased from 39 to 76 for G1-OP and from 24 to 72 for G2-90D. Overall success rates rose from 17.1% to 44.1% for G1-OP and from 13.4% to 31.1% for G2-90D. The company also said the data composition included 39,000 hours of “instruction-video” data, 32,000 hours of robot trajectory data and 30,000 hours of complete “instruction-video-action” data, used in stages.

For Japanese companies, the competitive requirement may be less about the performance of an individual robot platform and more about building a training foundation that combines video, trajectories and failure data. With transfer across different platforms and zero-shot evaluation taking center stage, there is a growing range of capabilities that are difficult to explain through control systems designed for a single model of robot.