High-Fidelity Real2Sim
The Real2Sim tool automatically converts captured real-world scenes and objects into simulation-ready digital assets with photorealism and physical consistency.
00:00
Simulator - DISCOVERSE
DISCOVERSE is the world’s first embodied AI end-to-end simulation framework based on 3D Gaussian Splatting (3DGS). It builds a complete Real2Sim2Real loop, enabling high-fidelity digitization of real-world scenes and zero-shot transfer of simulation-trained policies to real robots. With highly compatible development interfaces, DISCOVERSE supports rendering of RGB, depth cameras, LiDAR, force sensing, IMU, and tactile sensors, as well as various robot platforms including wheeled, legged, and desktop robots. DISCOVERSE provides a new infrastructure for embodied AI, combining high fidelity, high throughput, and high usability.
To bring robots to every home, DISCOVERSE supports efficient online Real2Sim2Real (RSR) training. Built around a simulation engine, it creates a complete RSR loop covering asset generation, scene rendering, model training, system evaluation, and Sim2Real transfer, addressing key challenges in embodied AI including data scarcity, high trial-and-error costs, and limited generalization. Real-world scenes, objects, and robot models are transformed into reusable high-fidelity digital assets through reconstruction, AIGC, or model import, enabling scalable generation of diverse 3D environments and interactive objects. The World Model simulates scene dynamics and robot-environment interaction. Through domain randomization and scene generation, it introduces variations in lighting, textures, and physical parameters to generate large-scale environments and improve model robustness. Sensor and actuator models simulate RGB, depth, LiDAR, IMU, force, and tactile feedback, while incorporating noise, latency, and control errors. Ground-truth data including poses, segmentation, contact forces, and joint states supports automated annotation, error analysis, and continuous optimization through evaluation feedback.
Key Features
High-Throughput Parallel Rendering
On consumer-grade GPUs, the physics engine supports multi-robot simulation with up to 1,000 DoF at 1,000 fps. The rendering engine achieves up to 10,000 fps at 640×480 resolution, supporting up to 2,048 parallel rendering streams.
End-to-End Model Validation
Building a complete loop from high-fidelity asset creation (Real2Sim), large-scale efficient training, to real-world deployment (Sim2Real), enabling large-scale robot learning and benchmarking of complex robotic systems.