fullseye

Fullseye 3DGS — how to use it (one command)

日本語 · English · 简体中文 · 繁體中文 · 한국어 · Deutsch

Turn a MuJoCo sim scene into a 3D Gaussian Splatting model and produce a GIF you can spin all the way around plus novel-view images. Camera poses come from the sim’s ground truth, so no COLMAP is needed.

The simplest way

In the imgevolve folder:

3dgs go2 --open

That alone turns go2 (a quadruped robot) into 3DGS and automatically opens the finished full-turn GIF.

• Change the scene: 3dgs cassie / 3dgs apollo / 3dgs anymal / 3dgs spot
• Your own MJCF: 3dgs <your local working path>\path\to\scene.xml
• See the list: 3dgs --list

Quality presets

3dgs go2 --quality fast       :: 128px / 8k Gaussians (a few seconds, for a quick look)
3dgs go2 --quality balanced   :: 256px / 20k Gaussians (default)
3dgs go2 --quality high       :: 384px / 45k Gaussians (the cleanest)

Cleaner still (densify)

3dgs go2 --quality high --densify --open

With --densify, training grows the number of Gaussians automatically to raise detail (native gsplat only). On go2 it grows from about 8k to about 50k, and the body and legs come out smoother. It finishes in a few to a dozen or so seconds.

The backend is automatic

• If native gsplat (tiled CUDA) is available it uses that automatically (fast and detailed, hundreds of it/s)
• Otherwise it falls back automatically to pure PyTorch (slower, but it works)
• You can also pick explicitly with --backend torch / --backend gsplat

The launcher sets up the environment (CUDA / compiler) automatically, so you don’t have to worry about vcvars and the like.

From Studio

Launch spikes/studio_app.py → in the “View sim model in 3D / make 3DGS” panel, pick the scene name (click a chip or type it) and the quality, then “3DGS train 🎇” → when it’s done the full-turn GIF opens.

Output

Under out/3dgs_<scene>/ (or wherever --out points):

• turntable.gif … full-turn preview
• novelview.png … left = ground truth / right = novel-view render
• gaussians.npz … trained Gaussians (npz)
• gaussians.ply … standard 3DGS .ply (with native). Drag and drop it into a web viewer such as SuperSplat to open it
• report.json … metrics such as PSNR

Requirements

• A GPU-training venv .venv-gsplat (torch cu128)
• To use native, .gsplat-cuda (CUDA 12.8) plus the C++ tools of VS BuildTools. Details and reproduction steps are in docs/GSPLAT_NATIVE_WINDOWS.md

Honest note: how much --densify helps is scene-dependent. A solid mass like go2 comes out clean, but a thin biped like cassie can overfit to the training views so the hold-out gets a little soft. The recommendation is to try without it first and add it if you want more.

Playing back motion (–motion)

3dgs go2 --motion --open

Instead of a still, this makes a 3DGS of a moving robot. How it works:

1. Train 3DGS in the canonical pose
2. Determine which MuJoCo body (link) each Gaussian comes from by segmentation and rig it
3. Move the joints (default = a sine wave), and for each frame do rigid-body skinning from the body poses (the sim’s ground truth) → re-render → motion.gif

Because a robot is a set of rigid links, it moves naturally without full 4D-GS. Change the number of frames with --frames N.

Honest: the default motion is a demo sine wave (not an actual walking policy). Since the sim’s poses are ground truth it never falls apart, but slight noise can appear near the feet (the boundary of body assignment for the initialization points).

Auto-generating a gait

3dgs go2 --motion --gait trot --open

--gait trot auto-generates and plays a quadruped trot gait (diagonal legs stepping in phase). It detects the legs automatically from the joint names (FL/FR/RL/RR or LF/RF/LH/RH + thigh/calf), so it works on go2 and anymal. Models it can’t detect fall back to a sine wave. To use the output of an actual walking policy, use --motion-file traj.npy (a qpos trajectory (F,nq)).

This site is open source. Improve this page.