An overview of how Genie Sim works. | Source: AGIBOT
AGIBOT Innovation Technology Co. today introduced Genie Sim 3.0, a next-generation robot simulation platform powered by NVIDIA Isaac Sim, at CES 2026. The company said it delivers a unified, open simulation workflow that brings together digital asset generation, scene generalization, data collection, automated evaluation, and physics-based simulation in a single toolchain.
AGIBOT said it designed Genie Sim’s core module, Genie Sim Benchmark, as a standardized evaluation system to establish an accurate and authoritative benchmarks for embodied intelligence.
Genie Sim 3.0 draws from more than 10,000 hours of synthetic datasets, including real-world robot operation scenarios. The system integrates 3D reconstruction with visual generation to create a high-fidelity simulation environment. It also uses large language model (LLM) technology to generate scenes and evaluation metrics.
The evaluation system covers more than 200 tasks across over 100,000 scenarios to build a comprehensive capability profile for models. AGIBOT said the platform can accelerate model development, reduce reliance on physical hardware, and empower innovation in embodied intelligence.
The Shanghai, China-based company was established in February 2023. AGIBOT offers a range of humanoid robots and mobile robots for cleaning and service settings.
In December, AGIBOT released its updated humanoid, AGIBOT A2, in December. Later that month, the company rolled out its 5,000th mass-produced humanoid robot at its factory.
Genie Sim 3.0 provides a full-stack pipeline for training and evaluation
How Genie Sim’s LLM-driven scene generation works. | Source: AGIBOT
AGIBOT said its integrated technologies enable it to deliver real-time, high-fidelity, and high-precision simulation environments.
The system captures real-world environments with Skyland Innovation’s MetaCam handheld 3D laser scanner. This scanner combines high-resolution RGB images, 360° lidar point clouds, and centimeter-level RTK (real-time kinematic) positioning. Interactable objects can be converted into simulation-ready assets from a single 60-second orbital video, significantly accelerating scene construction, said the company.
A defining feature of Genie Sim 3.0 is its LLM-driven natural-language scene generation, AGIBOT claimed. Users can describe environments conversationally, and the system automatically produces structured scenes, visual previews, and thousands of semantic variations without manual logic coding. Vision-language models further refine and tune these scenes to meet specification-level needs, enabling rapid adaptation and strong model generalization, it claimed.
In addition, Genie Sim 3.0 integrates real industrial-scene datasets into both training and evaluation pipelines. This approach significantly shortens algorithm-validation cycles and reduces dependence on hardware infrastructure.
AGIBOT provides over 10,000 hours of synthetic data
An overview of how Genie Sim handles multi-dimensional evaluation. | Source: AGIBOT
AGIBOT has also released over 10,000 hours of open-source synthetic data covering more than 200 tasks with multi-sensor modalities such as RGB-D, stereo vision, and whole-body kinematics. An intelligent data-collection toolkit supports both low-latency teleoperation and automated task programming. With auto-annotation and a recovery mechanism that resumes collection after task failures, it can reduce the cost and time of dataset production.
For evaluation, Genie Sim 3.0 provides more than 100,000 simulation scenarios, which AGIBOT said moves beyond single-metric benchmarks to construct full-spectrum capability profiles for embodied models. Using LLMs, the platform can auto-generate executable evaluation workflows across diverse semantic, spatial-reasoning, and manipulation dimensions. The company said it outlines model strengths, limitations, and optimization pathways.
All simulation assets, datasets, and automated evaluation source code are fully open-source. Developers and researchers can access AGIBOT Genie Sim on GitHub.
Tell Us What You Think!