Assets

Panoramic multi-agent system header

The camera rotates around the measured COLMAP point cloud, recorded three-UAV trajectories, and semantic observation anchors. This is not a synchronized flight replay. Labels mark supporting UAV observation poses, not reconstructed object centers; the COLMAP reconstruction has arbitrary scale, not calibrated meters.

The display uses UAV1 from Pose 50 onward, excludes the isolated UAV2 Pose 105, and retains UAV3 unchanged (56, 175, and 123 trajectory points). The 58 semantic labels share exactly the same display mask. These are visualization-only filters; raw observations and the QA benchmark evidence are unchanged. The map is enlarged 1.4×, and the only footer reads Panoramic multi-agent system (32 px at 1080p).

The website MP4 is a compressed 720p version of the original 24-second render, retaining its frame rate and playback duration.

Captioned demos

Both captioned animations loop automatically and are 15 seconds long at 10 FPS. Town04 uses a 2 × 2 view layout; Town05 uses a 5 × 2 view layout. Both are served as compressed MP4 videos.

Robot Dog demo

Robot Dog video: 15.9 seconds, 1280 × 720, 30 FPS. The robot navigates to a basketball court, with first-person, third-person, and point-cloud views. Playback is muted by default; the native controls can enable the original audio.

MemNTN paper figures

Figures from Memory-Native Non-Terrestrial Networks for Embodied Intelligence, IEEE Communications Standards Magazine, 2026.

Figure 1 geographic data sources: OpenStreetMap, LEOPath, and Mega-NeRF/Mill-19.