AgiBot unveils foundation model GO-3 Preview, leads three embodied AI benchmarks

AgiBot unveiled its embodied AI foundation model, GO-3 Preview, on November 3, 2026. The company said it ranked first in three benchmarks: RoboDojo, RoboTwin and RoboColiseum. The model is designed to link pretraining, post-training and learning after deployment on physical robots. AgiBot plans to make it fully open source by the end of 2026.

GO-3 Preview presents not only benchmark results but also a post-training setup that combines 100 physical robots with simulation and world models. For those evaluating robot learning systems, rankings and real-world task records are distinct measures; whether learning methods and evaluation conditions can be examined after release will be a concrete factor in assessing the system.

The first-place results in three benchmarks and the post-training environment using 100 physical robots demonstrate different capabilities. The former are comparative results on specific benchmarks, while the latter is described as a system for feeding physical experience back into learning. The novelty here, therefore, lies not only in the number of top rankings but also in presenting WRAM for pretraining, Genie Evolver for post-training in physical and virtual environments, and learning after deployment as parts of a continuous design. However, benchmark rankings in the announcement materials alone do not establish generality or task reliability in real-world settings. In the system's design, WRAM links task understanding to predictions of future states and action generation, while Genie Evolver uses physical trajectories, simulation and alternative action choices in world models. The approach treats model learning and robot execution as connected, using experience from execution to improve the next policy. The figure of 100, however, refers to the scale of physical robots incorporated into the learning system; it does not indicate the number of robots AgiBot has deployed or put into commercial use. The announcement materials do not specify the conditions or metrics behind the benchmark rankings, or the success rates and task times achieved by the robots. In related coverage, this newspaper reported on a proposal for AgiBot's Worken robot-operation metric. It also reported on AgiBot's 24-hour store livestream, which included figures for completed tasks. These concern a different company, but offer a point of comparison in showing why benchmark assessments of models should be distinguished from measuring what was done, and how much, in real-world settings. For GO-3, too, evaluation will depend on the reproducibility of its rankings, how data from physical robots is collected and selected, how failed experiences are handled, and how widely experience gained after deployment is shared. AgiBot said it plans to make the model fully open source by the end of 2026, but the materials do not specify what will be released, the license or the terms for using the training data. Key questions include how much of the model and training process can be reproduced after release, whether capabilities transfer to tasks beyond the three benchmarks and to different robots, and how safety on physical robots will be evaluated. For now, it is important not to equate the company's plans and reported rankings with operating performance independently verified by third parties.

GO-3 Preview is based on the company's proprietary embodied AI architecture, WRAM (World Reasoning Action Model), which combines task understanding, prediction of future states and action generation in a single system. The company said pretraining uses human first-person videos and robot motion trajectories as training data. The post-training system, Genie Evolver, combines 100 physical robots with physics simulation and world models. It also uses CEL (Collective Experience Learning) to aggregate experience from physical execution. Together with the agent platform Genie Master, it forms a learning system that includes deployment on physical robots.