worldarena-track1-top20-trajmaps-v0

# WorldArena 2.0 Track 1 Top-20 Trajectory Maps v0 > **Public derivative dataset — Apache License 2.0.** This repository is a > cleaning & trajectory-visualization **derivative** produced from the public > **WorldArena 2.0** dataset. It is **not** an official WorldArena release, and > WorldArena does **not** endorse it. Upstream license (Apache-2.0), name, source, > copyright and license notices are retained. See §10. ## 1. What this is **Data name:** WorldArena 2.0 Track 1 Top-20 Trajectory Maps **v0** This repository contains **preprocessed model-input conditions** derived from the **official WorldArena 2.0 Track 1 test set, episodes 1–20**. It is **NOT** the official ground-truth video, and it is **NOT** a submission artifact by itself. ## 2. Source - Source episodes: `episode1` … `episode20` from the official WorldArena 2.0 Track 1 test set. - This repo lives at remote path `worldarena_track1/top20_v0/`. ## 3. Conversion method (how the conditions were produced) For each episode: 1. Read the official HDF5 (action / robot states). 2. Apply **aloha-agilex URDF forward kinematics**. 3. Use a **fixed head-camera calibration**. 4. Draw the **left & right arm skeletons**, **end-effector points**, and **XYZ axes**. 5. Render the **full T-frame** trajectory map video `traj_maps_full.mp4`. 6. Keep the raw control deltas as `delta_action_full.npy` with shape **(T−1, 14)**. 7. Additionally slice **C4-continuous windows** of **49 frames** with **stride = 48** (`windows/window_xxx/traj_maps.mp4`). 8. When the last window is shorter than 49 control frames, the **final control frame is repeated** to pad it to length. ## 4. Current validation status | Item | Status | |------|--------| | Full trajectory generation | **validated** | | Window preprocessing (C4, 49f, stride 48) | **validated** | | Autoregressive (cross-window) inference | **pending AB GPU validation** | ## 5. FPS notes | Quantity | Value | |----------|-------| | Source action Hz | `unknown` | | Control full-video container FPS | **24** | | Model window FPS | **16** | | Submission FPS | **24** | ## 6. Directory structure ``` <repo root>/ ├── README.md # this file ├── MANIFEST.sha256 # SHA-256 of every data file (repo-root relative) ├── inventory.json # full read-only audit of the source delivery ├── upload_file_list.jsonl # upload work-order (local→remote map + sha256) ├── download_top20.sh # teammate download helper ├── verify_manifest.py # recompute & compare SHA-256 after download ├── TEAMMATE_DOWNLOAD.md # teammate instructions └── worldarena_track1/top20_v0/ # the dataset content (local hierarchy preserved) ├── INDEX.md ├── delivery_manifest.jsonl ├── selected_top20.jsonl ├── failures_top20.jsonl # empty == 0 failures (success) ├── qa_report_top20.json ├── report_top20.md ├── episode1/ │ ├── caption.txt caption_1.txt caption_2.txt │ ├── metadata.json checks.json projection_sanity.json │ ├── frame.png frame_source.png overlay_frame0.png │ ├── delta_action_full.npy # (T-1, 14) │ ├── traj_maps_full.mp4 # full T-frame map │ ├── window_manifest.json │ └── windows/window_xxx/traj_maps.mp4 # C4 49-frame windows ├── episode2/ … episode20/ └── ... ``` ## 7. How to download (teammates) ```bash # 1) install the lightweight hub client (no heavy DL deps) python -m pip install modelscope-hub # 2) set YOUR OWN token (do not use anyone else's) export MODELSCOPE_API_TOKEN=ms-xxxxxxxxxxxxxxxx # 3) download (this is a PUBLIC dataset; a token is optional but recommended) ms-hub download "yilusun/worldarena-track1-top20-trajmaps-v0" \ --repo-type dataset \ --revision master \ --local-dir ./worldarena_track1_top20 # or use the helper script in this repo: bash download_top20.sh ``` ## 8. How to verify integrity (SHA-256) ```bash python verify_manifest.py --root ./worldarena_track1_top20 --manifest ./worldarena_track1_top20/MANIFEST.sha256 ``` Every data file's SHA-256 is recorded in `MANIFEST.sha256` (paths relative to the repo root). `verify_manifest.py` recomputes each hash and reports any mismatch. ## 9. Important caveats for AB inference - **Do NOT reuse the official `first_frame` for every window.** - `window0` uses the **official first frame**. - Subsequent windows (`window1`, `window2`, …) should use the **predicted last frame of the previous window** (autoregressive chaining). - `delta_action` currently does **not** enter the control V2V inference. - **Cross-window autoregressive quality is still pending GPU validation.** ## 10. License & attribution - This repository is a **derivative cleaning & trajectory-visualization product** generated from the **public WorldArena 2.0** dataset (Track 1 test set, episodes 1–20). - **Upstream dataset license: Apache License 2.0.** This repository is distributed under the same **Apache License 2.0**. - This repository is **NOT** the official WorldArena release repository. - The upstream project **name, source, copyright notices, and license** are retained and respected here. - Nothing in this repository implies **official endorsement** by, or affiliation with, the WorldArena project or its authors. - This repository does **not** redistribute the raw official HDF5 or official archives — only derived trajectory-map videos, delta-action arrays, frames, captions, and metadata.