habitat-gs: 3D generator — what it does and what you need to run it

Habitat-GS is an open source 3D generator from ZJU3DV. It is designed to create highly detailed scenes and dynamic avatars based on Gaussian splatting. The tool's main goal is to provide an environment for robot navigation and embodied AI, where an agent interacts with realistic geometry and moving objects.

What it does

As described by the authors, Habitat-GS recognizes scene assets by file name suffix: a scene must end in .gs.ply or .3dgs.ply. This indicates support for the ply export format.

The tool supports avatar rendering using explicit Gaussians and CUDA LBS. According to the developers, no neural forward pass is used at runtime. The generated driver.pkl file contains precomputed joint matrices for rendering, as well as proxy capsules that are used to handle dynamic obstacles at the NavMesh level. As described by the authors, this guarantees that the agent cannot pass through Gaussian avatars.

The repository is tagged: computer-vision, eccv2026, embodied-ai, embodied-navigation, gaussian-avatar, gaussian-splatting, navigation, robotics, simulator, vln, vln-ce. The authors describe the project as a high-fidelity navigation simulator with dynamic Gaussian splatting, presented at ECCV 2026.

What you need to run it

The platform, as described by the author, is CUDA. This follows from the mention of CUDA LBS in the avatar rendering description.

The VRAM requirement, as described by the author, is 80 GB per GPU for standard full fine-tuning with a DeepSpeed ZeRO-2 configuration. The authors recommend an A100 80GB. For consumer GPUs such as the RTX 3090 or 4090, a --lora flag is offered, which enables LoRA training with lower memory consumption.

The code is written in C++. The repository was created on April 14, 2026, and the last code change was on September 15, 2026.

The code license is Project Registration License v1.0. The license terms permit use, copying, modification and distribution of the software, including for commercial purposes. For personal, educational and research use, registration is not required. For use in projects by companies and organizations, free prior registration via a form is required; permission is granted automatically. Trademarks and patents are not licensed.

Who it suits

Habitat-GS will suit research groups and labs working on embodied navigation, robotics and agent training in realistic environments. The presence of dynamic Gaussian avatars and obstacle handling makes it applicable to tasks where moving objects and interaction with them matter.

The tool is unlikely to suit those looking for a lightweight static 3D generator for quick export to game engines or for running on an ordinary laptop. The 80 GB VRAM requirement for full fine-tuning limits use without access to server GPUs. LoRA mode lowers the threshold, but still assumes a discrete graphics card at the RTX 3090 or 4090 level.

Habitat-GS is a specialized tool for navigation simulation with dynamic Gaussian splatting. Its strength is the combination of avatar rendering and obstacle handling in a single pipeline. At the same time, the high VRAM requirements for full training make it primarily a choice for teams with access to server accelerators.