Demonstration gallery
The framework README uses one continuous, silent gallery of 44 native paper and hardware clips. A camera visits five groups, then pulls back to the whole wall before returning to the first group. Wider gutters separate the groups. There are no added titles, captions, logos or labels inside the animation.
| Order | Group | Native footage |
|---|---|---|
| 1 | MDOC | Four single-robot scenes and four multi-robot CBS scenes from the original MDOC project. |
| 2 | MD-COAS | Two planar scenes (including the paper’s L10, seed 8 example) and 7-DoF avoidance, each pairing diffusion with 20 candidate trajectories, beside the D3IL execution environment. |
| 3 | 2GO | Two quadruped stepping-stone scenes and four humanoid corridor variants. |
| 4 | MGA | Twelve surface geometry/material variants, three peg-insertion variants and three humanoid contact tasks. |
| 5 | Hardware | 2GO humanoid execution, plus MGA rigid/curved/compliant scanning and peg insertion. |
Website playback
This website presents individual clips rather than the README’s moving gallery. All 2GO animations retain their original frame timing, and hardware excerpts retain the original source-video timing. They loop silently without added acceleration. The homepage hero combines the original MGA contact and 2GO locomotion simulations with real G1 hardware using AR virtual obstacles and MGA hardware surface interaction and insertion. Three additional Unitree H1 contact clips complete the hero’s 3 × 3 grid. The gallery retains all six original research demonstrations and adds three hardware clips. An additional row shows MGA Unitree H1 fixed-stance pushing, contact-force regulation, and unjamming from the paper supplement. These humanoid clips preserve the original GIF timing, including the final 800 ms hold; they preview research beyond the currently qualified V1 recipes.
| Website clip | Preserved source interval |
|---|---|
| 2GO Go2 stepping stones | Full original GIF, 16.28 seconds |
| 2GO G1 corridor | Full original GIF, 34.56 seconds |
| 2GO real G1 | 92–112 seconds of the original hardware video, played over 20 seconds |
| MGA hardware excerpts | 12–24 seconds of the original deployment video, played over 12 seconds |
| MGA humanoid pushing | Full original GIFs: fixed stance 1.54 s, force regulation 2.80 s, unjamming 2.12 s |
Other website planning and simulation clips retain presentation timing. Source animation timing does not establish planning latency or synchronize separate experiments. See each result’s playback context.
README gallery files and regeneration
showcase.gifis the GitHub-compatible looping animation.showcase.mp4is the 1344 × 992 video with the same 24-second camera journey.showcase-poster.pngshows the complete gallery in one frame.architecture.svgis the separately editable framework diagram.
The compact source clips are checked in under showcase_sources/v2/.
manifest.json records all 44 checksums,
original source identities, crop bounds, selected time ranges and playback
normalization. It uses symbolic source roots rather than personal machine paths.
The gallery can be rebuilt without a simulator or the original paper folders:
python -m pip install Pillow numpy imageio-ffmpeg
python scripts/visualizations/build_readme_showcase.py
Within MD-COAS, each column pairs the diffusion process above with the generated
candidate trajectories below. The planar views use L6/seed 0 and L10/seed 8.
The arm view uses L1/seed 0 and shows the TCP projection of all 20 planned
7-DoF trajectories, sourced from trajectory_modes_plan.gif rather than the
execution trace. Candidate animations reveal the paths over their horizon;
they are separate runs from the adjacent D3IL execution footage.
To inspect the composition before encoding:
python scripts/visualizations/build_readme_showcase.py \
--preview-only --output /tmp/generative-gallery-preview
The separate prepare_readme_media.py script recreates the compact inputs when
original source folders are available. Its --help documents the source-root
overrides. It reads those folders and writes only to the requested output path.
Presentation and release boundaries
All robot motion, diffusion trajectories and scene geometry come from the original footage. Colors are preserved. Crops remove source titles, subtitles, rulers, plots and empty margins; frames retain their aspect ratios. No robot motion or scientific data is synthesized. In the README gallery, clips loop and their selected time ranges are rescaled to six seconds for simulation/planning or eight seconds for hardware. Playback and the moving gallery camera do not indicate execution latency, synchronization between experiments or real-time performance.
The gallery presents the broader research behind the framework. Its contents are not a device-support or reproduction matrix:
- The original MDOC project includes CBS multi-robot coordination. This V1 repository exposes the single-robot planning component; the multi-robot application remains in the MDOC project.
- MGA's humanoid contact tasks are paper demonstrations. Humanoid walking/push and MGA GPU qualification remain follow-up work for this release.
- The MGA
soft_*footage shows a rigid robot arm contacting compliant material. It does not introduce the excluded soft-robot or co-design systems. - Hardware footage demonstrates the research systems used for the papers; installing V1 alone does not establish a qualified driver/controller setup for every robot shown.
Use the recipe catalogue, compatibility matrix and V1 evidence for runnable configurations and validated paths.