AgiBot updates simulation platform ‘Genie Sim,’ adding support for flexible-object simulation

AgiBot (Shanghai) Technology Co., Ltd. updated its simulation platform Genie Sim on October 28, 2026. The update strengthens capabilities for handling the deformation and contact of flexible objects and integrates scene reconstruction, data collection, automated evaluation and model post-training into a single toolchain. In an experiment evaluating models trained on real robots in simulation without adjustment, the company said simulation and real-world task scores had a 96.7% correlation across four robot embodiments and 22 tasks.

AgiBot says the platform allows users to evaluate problems identified in simulation and conduct post-training in the same system, handling high-cost, high-risk tasks in a simulated environment. Whether it can reduce the burden of real-world testing for tasks involving flexible objects depends on the scope of the correlation figure reported by the company and how well results can be reproduced on real robots.

What stands out in the update is not only the simulation capabilities for flexible objects, but also the integration of comparative evaluation against real robots and subsequent post-training on the same platform. Flexible objects change shape and contact conditions during manipulation, making the gap between simulation and real-world performance a challenge for simulations centered on rigid bodies. AgiBot described an approach that combines adjustments to physical properties using real-world data with scene estimation by a world model. However, the 96.7% figure represents the correlation in scores across four embodiments and 22 tasks; it does not show that simulation results match real-world success rates across all tasks. The scope of the evaluation should be kept in mind. Our article, AgiBot to unveil seven new products at AI WAVE; G2 to undergo 24-hour testing at convenience stores starting Nov. 2, reported the company’s plan to announce seven items in succession, including data, a task platform and robot embodiments, and described the role of full-stack collaboration. Genie Sim is the training and validation platform within that plan, connecting real-world data to simulation, evaluation and post-training. While our article published the same day, AgiBot unveils embodied-agent framework ‘Genie Master,’ tests autonomous improvement in real stores, covered real-store execution and feedback on results, Genie Sim can be understood as handling validation and iterative training in simulated environments. The announcement does not confirm a technical connection between the two. The broader industry implication is that simulation fidelity and evaluation methods are being incorporated into model-improvement workflows alongside robot movement. The four-axis benchmark is designed to measure more than a single task-success rate, including whole-body coordination and skill transfer between embodiments, and could serve as a basis for comparing models deployed across multiple embodiments. However, the correlation figure reported by the company alone does not establish the breakdown by task, variation across embodiments or reproducibility in real-world settings. Key points to watch include the 22 tasks, evaluation procedures, the scope and terms of use for simulation assets and datasets, and real-world performance after post-training.

The company’s physics engine, Genie Sim Engine, uses multiple physics solvers that separately handle contact between rigid and flexible objects, and manages dynamic feedback and state consistency. The company says it uses real-robot operation data to adjust properties of objects in simulation, including mass density, friction coefficients and damping, and reports a 20-fold increase in processing speed. Genie Sim Dream, which uses a world model, estimates how a scene will change after an action based on real-world video, and supports VLM-based judgments of task completion and identification of failure points. The Genie Sim Benchmark added four evaluation dimensions: dexterous manipulation, whole-body coordination, precision manipulation and skill transfer between embodiments. The company also offers Genie Sim × RLinf, a parallel post-training framework that selects tasks based on evaluation results and conducts reinforcement learning across multiple simulation environments. Genie Sim Sandbox, an open-world simulation environment, covers homes, factories, supermarkets, restaurants and offices. The company says it will update the simulation code on GitHub and publish simulation assets and related datasets on Hugging Face and ModelScope.