GS-Playground

A high-throughput simulator from Tsinghua and others that uses batched 3D Gaussian Splatting to render photorealistic images quickly.

GS-Playground was proposed by Tsinghua University together with Motphys, Dexmal, and more than a dozen other organizations; its paper was released in April 2026, its project page shows acceptance at RSS 2026, and its code is open-sourced on GitHub. It addresses a tension in vision-based robot learning: traditional simulators don't look realistic, while photorealistic rendering is too slow to run at the scale needed for visual reinforcement learning. Its approach combines a self-developed parallel physics engine, MotrixSim (compatible with the MJCF format), with batched 3D Gaussian Splatting (representing a scene with a large number of colored ellipsoids, which renders quickly), reaching a total rendering throughput of about 10,000 frames per second across 2,048 parallel environments at 640×480 resolution, and it can automatically generate a simulatable digital twin from a single RGB photo. The paper validates the approach on quadruped and humanoid locomotion, visual navigation, and robot-arm grasping, and deploys it to real robots including the Unitree Go2 and G1.

ExampleThe project page states that generating a scene asset ready to drop into simulation from one RGB photo takes under 5 minutes, and that humanoid physics simulation runs 32 times faster than MuJoCo.

Also called

A High-Throughput Photorealistic Simulator for Vision-Informed Robot Learning

Related

Gaussian Splatting-based Simulation · 3D Gaussian Splatting · Real-to-Sim · Digital Twin · Batched Rendering · DISCOVERSE

Sources

GS-Playground (arXiv 2604.25459)
GS-Playground project page

As of

2026-09