On the first day of CES 2026, Chinese robotics company AGIBOT (智元机器人) officially released Genie Sim 3.0, the world’s first open-source simulation platform driven by large language models (LLMs), and simultaneously open-sourced a dataset containing tens of thousands of hours of real robot operation scenarios, making it one of the event’s key technical highlights.
Built on NVIDIA Isaac Sim, Genie Sim 3.0 integrates 3D reconstruction and visual generation to create digital twin–level, high-fidelity environments. It pioneers an LLM-driven approach that enables the generation of tens of thousands of scenarios within minutes. At the same time, it open-sources a simulation dataset comprising over ten thousand hours of real robot operation scenarios, and establishes a multi-dimensional intelligent evaluation system covering more than 100,000 scenarios, providing a comprehensive profile of model capabilities.
By offering a full closed-loop solution covering digital asset generation, scenario generalisation, data collection and automated evaluation, the platform is expected to significantly accelerate model training and validation, reduce reliance on physical hardware, improve R&D efficiency for developers and researchers, and promote innovative applications in embodied intelligence.
Project homepage: https://agibot-world.com/genie-sim
GitHub: https://github.com/AgibotTech/genie_sim
ModelScope: https://modelscope.cn/datasets/agibot_world/GenieSim3.0-Dataset
Genie Sim 3.0 introduces a new era of embodied simulation through five core highlights.
First, a digital twin–level high-fidelity simulation environment integrates 3D reconstruction, visual generation technologies and physics engines. Using tools such as the MetaCam handheld 3D laser scanner, it achieves millimetre-level replication of real environments. Visual generation models enhance realism, enabling accurate mesh simulation models to be generated from a 60-second surround video, balancing both visual and physical fidelity.
Second, natural language–driven scenario generation and generalisation allows users to automatically generate and generalise thousands of training and testing scenarios within minutes simply by inputting natural language instructions, without manual logic coding. Generated scenarios support conversational editing, enabling flexible addition or removal of details and layout adjustments with high efficiency.
Third, fully open-source datasets and efficient data collection provide large-scale simulation datasets for the embodied intelligence domain, covering more than 200 tasks and over ten thousand hours, with multi-sensor information and multiple generalisation dimensions. Dual-mode data collection tools and automated annotation features are offered, alongside an original error recovery mechanism that enables zero-shot Sim2Real transfer, achieving task success rates exceeding those trained on real-world data.
Fourth, a comprehensive evaluation system covering over 100,000 scenarios establishes an all-round benchmark. Using LLMs to automatically generate evaluation instructions and workflows, and VLM technology to assess models across dimensions such as semantic understanding and spatial reasoning, the system keeps the discrepancy between simulation and real-world evaluation below 10%, enabling efficient validation without physical robots.
Fifth, deep simulation of real operational scenarios integrates industrial use cases such as supermarket restocking, logistics sorting and power inspection, faithfully reproducing real working environments. It supports generating training datasets and end-to-end evaluation systems based on reconstructed assets, reducing R&D costs and validation cycles, and enabling “zero hardware deployment with fully realistic verification.”
Genie Sim 3.0 fully open-sources its core code, datasets and digital assets, making them available to researchers, engineers and industry experts worldwide. By providing abundant simulation scenarios, high-fidelity training environments and scientific evaluation standards, the platform aims to allow users to focus on algorithm and model innovation. Developers can access related resources via the GitHub open-source page at https://github.com/AgibotTech/genie_sim.