Today, Zhiyuan Robotics officially announced the launch of its independently developed large-scale simulation framework, AgiBot Digital World. This framework provides flexible simulation data generation solutions for robot operation, pre-trained large-scale simulation data, and unified model evaluation standards. A massive amount of simulation data is also being open-sourced simultaneously.
AgiBot Digital World is a simulation framework specifically designed for research and application of robot manipulation skills. It integrates a massive amount of realistic 3D assets, diverse expert trajectory generation mechanisms, and comprehensive model evaluation tools. Through high-fidelity simulated training scenarios and fully automated data generation, it achieves a series of embodied simulation services.
This framework leverages a large-scale, high-precision 3D asset library built with a wide variety of objects, scenes, and robot models. Combining realistic visual rendering, accurate physical simulation, and automated task and scene generation, it can vividly simulate various robot training scenarios. Its multiple expert trajectory generation strategies, coupled with domain randomization and data augmentation techniques, can automatically generate diverse and robust large-scale expert trajectory data. A complete evaluation process has also been established, encompassing expert trajectory generation, model training, benchmark testing, and deployment verification.
Furthermore, Zhiyuan Robotics is open-sourcing a massive simulation dataset, the AgiBot Digital World Dataset. This dataset covers 5 major scene categories, 180+ specific item categories, 9 common materials, and 12 core skills. It is characterized by high quality, fast generalization, diverse tasks, and flexible application. This is Zhiyuan Robotics' second open-source initiative in embodied intelligence data, following the open-sourcing of the million-scale real-world dataset project, AgiBot World, last December.