AGIBOT updates Genie Sim with deformable-object simulation, reports correlation with real-world performance

leaderobot.com reported on October 28, 2026, that AGIBOT had substantially updated its simulation platform, Genie Sim. According to the article, the update focuses on simulating deformation and contact involving deformable objects such as cloth and strings. In an evaluation using four types of robot platforms and 22 complex tasks, the correlation between simulation and real-world task scores was 96.7%, the report said.

According to leaderobot.com, Genie Sim handles manipulation of deformable objects and showed a 96.7% correlation between simulation and real-world scores across four types of robot platforms and 22 tasks. The figure can serve as an indicator when connecting simulation results to real-world evaluation in robot development, but it applies only to the evaluation conditions described in the article.

The report centers on an update to a development platform that reproduces manipulation of soft objects in a simulation environment and links evaluation with real-world robots, rather than on new capabilities in the robot platforms themselves. It identifies changes in the shape and contact of objects such as cloth and strings during manipulation as a challenge, and says Genie Sim has strengthened its capabilities for handling deformable objects and complex contact. The reported 96.7% is described as the correlation between task scores for four types of robots and 22 tasks. It does not necessarily mean that simulation and real-world robots match across all operations or achieve the same success rate on each task. Interpreting the figure requires distinguishing the method used to calculate the correlation from the variation among individual tasks. According to the report, the update brings scene reconstruction, data collection, automated evaluation, model post-training and open-world exploration together in a single toolchain. This can be understood as a configuration that connects simulation not only to visual reproduction but also to the generation of training data, evaluation and further training. However, the article's description alone does not establish the conditions under which these capabilities are available or the degree to which each step is automated. The explanation that the system combines an engine and a world model indicates a design that uses physical contact calculations alongside predictions of scenes after actions, but the report does not say how much either contributed to the evaluation results. In a story, we covered AGIBOT’s release of the Physical Reasoning inference-training dataset, which addressed data for changing actions according to circumstances. The simulation evaluation described in this article concerns a different issue: training materials versus evaluation between simulation and real-world robots. There is no basis for concluding that the particular dataset was used in this evaluation or improved its performance. The evaluation also differs from the ranking covered in leaderobot.com’s report on an embodied intelligence evaluation ranking led by the University of Hong Kong. A position in a ranking, correlation between simulation and real-world task scores, and sustained operation in the field are separate metrics and cannot substitute for one another. To assess the implications for the industry, the focus should be not only on the single aggregate figure of 96.7% but also on results by robot platform and task, evaluation conditions, and reproducibility by third parties. The article does not provide details on these points, operational results in real-world settings, or the effect of simulation-generated data on real-world robot training and performance. Therefore, what can be confirmed at this stage is the report that AGIBOT updated a development toolchain that includes simulation of deformable objects, and its account of a correlation figure under limited evaluation conditions. The evaluation procedure and task-by-task results, the availability conditions for the features, and the scope of real-world verification will need to be examined.