Skip to content

SHOWCASE.md

Latest commit

 

History

History
106 lines (79 loc) · 3.09 KB

SHOWCASE.md

File metadata and controls

106 lines (79 loc) · 3.09 KB
Raw

MuGS Showcase Materials

This directory contains scripts and examples for generating demonstration materials.

Generated Materials

AndroidTwin × G1 hybrid (docs/images/showcase/androidtwin_g1/)

Unitree G1 humanoid in the INRIA kitchen scene — pipeline panel, static-vs-tracked bg comparison, and 16-frame animation. Full write-up in the project README.

Static Showcase Image

File: mugs_showcase.jpg

Three-panel comparison showing:

  • Left: MuJoCo simulation (physics-accurate but unrealistic visuals)
  • Middle: 3D Gaussian Splatting (photorealistic but no physics)
  • Right: MuGS Hybrid (best of both worlds)

Use for: Papers, presentations, README

Animated Demo

File: mugs_demo.gif

8-frame animation showing robot pick-and-place motion with MuGS rendering.

  • Loop: forward + reverse playback
  • Frame rate: 500ms per frame
  • Demonstrates real-time hybrid rendering

Generating Showcase Materials

Quick Generation

cd /home/ununtu/metabot-workspace/mugs
TORCH_CUDA_ARCH_LIST="8.6" python examples/yam_standalone_demo.py

This creates individual component images in outputs/yam_standalone_demo/.

Custom Showcase Image

from mugs.sensors import GaussianSensor, GaussianSensorConfig

# Render with return_components=True
result = sensor.render(model, data, camera_name, return_components=True)

# Access components
mujoco_only = result['foreground']   # Physics simulation
background = result['background']    # 3DGS photorealistic
hybrid = result['rgb']              # MuGS hybrid
mask = result['mask']               # Blending mask

Custom Animation

See examples/yam_wrist_camera_demo.py for keyframe animation example.

Key steps:

  1. Define motion keyframes (joint positions)
  2. Render each frame with MuGS
  3. Collect frames and save as GIF with PIL

Camera Viewpoints

Task Observation Camera (Wrist-Mounted)

  • Camera: robot/camera_d405 on YAM link_6
  • View: First-person gripper perspective
  • Use: Matches actual RL task observations
  • Demo: examples/yam_wrist_camera_demo.py

Showcase Camera (Side View)

  • Camera: Fixed third-person view
  • View: Shows full scene and robot motion
  • Use: Better visual impact for demonstrations
  • Demo: examples/yam_standalone_demo.py

Tips

High-Quality Rendering

  • Use 640×480 or higher resolution
  • Add good lighting in scene XML
  • Set headlight ambient="0.5 0.5 0.5" for visibility

CUDA Compilation

RTX 4090 with CUDA 11.6 requires:

TORCH_CUDA_ARCH_LIST="8.6" python script.py

Performance

  • Hybrid rendering: ~5000 FPS (640×480)
  • Background caching enabled by default
  • First render loads PLY (~2 seconds), subsequent renders are fast

Examples

Paper Figure

Use mugs_showcase.jpg - clean three-panel comparison with annotations.

Video/GIF

Use mugs_demo.gif or generate custom animation showing:

  • Robot manipulation tasks
  • Multiple camera angles
  • Different objects/scenes

Interactive Demo

Use yam_standalone_demo.py as starting point for real-time demos.