DISCOVERSE

An open-source real-to-sim-to-real simulation framework combining 3D Gaussian Splatting rendering with MuJoCo physics.

DISCOVERSE is an open-source robot simulation framework released jointly by Tsinghua University, Zhejiang University, and other institutions together with DISCOVER Robotics and D-Robotics (地瓜机器人); its paper was accepted at IROS 2025, and the code is MIT-licensed. It uses 3D Gaussian Splatting (3DGS, a method that reconstructs and quickly renders a real scene using a large number of colored 3D Gaussian points) for visuals and MuJoCo for physics: a real scene is first reconstructed with photorealistic appearance and then paired with collision and physics models, producing a simulated environment that “looks like the real world,” narrowing the visual sim-to-real gap. It supports parallel simulation of multiple cameras and sensors, is compatible with existing 3D assets, robot models, and ROS plugins, and has been adapted to embodiments including Airbot Play, AgileX PiPER, UR5e, Franka Panda, and the LEAP Hand. Its project page reports rendering 5 channels of 640×480 RGB-D cameras at up to 650 FPS (about 240 FPS on a laptop); the paper's imitation-learning experiments show better zero-shot sim-to-real transfer than other simulators.

ExampleA lab tabletop is scanned and reconstructed as a 3DGS scene inside DISCOVERSE; demonstrations are collected in simulation to train a grasping policy, which is then deployed directly to the same real tabletop with no real-robot fine-tuning.

Also called

Efficient Robot Simulation in Complex High-Fidelity Environments

Related

Gaussian Splatting-based Simulation · 3D Gaussian Splatting · MuJoCo (Multi-Joint dynamics with Contact) · Real-to-Sim-to-Real · Sim-to-Real Gap (Reality Gap) · Photorealistic Rendering

Sources

DISCOVERSE: Efficient Robot Simulation in Complex High-Fidelity Environments (arXiv 2507.21981)
DISCOVERSE 项目主页 (Chinese)
DISCOVERSE GitHub

As of

2025-07