| # GAUGE world-model videos and full trajectories, v1 |
|
|
| This local release packages the documented, completed shared-seed world-model experiments and the nine selected Genie 3 videos. It contains original MP4 videos and byte-for-byte copies of the **full raw trajectory JSON files**, before task-specific analysis cropping. It is prepared for addition under `world_models/v1/` in `InternRobotics/GAUGE-Dataset`; no remote upload has been performed by the build. |
|
|
| ## Scope |
|
|
| There are **3,249 video/trajectory pairs** across five task families and nine task/material settings. The three local models contribute 100 seeds per task with the physical negative prompt and ten seeds per task without it. Wan2.7 contributes ten seeds per task in each condition; Seedance 2.0 contributes ten seeds without it. Genie 3 contributes one selected `_1.mp4` per task, with unknown generation seed. |
|
|
| | Model | Negative prompt | Tasks | Videos | Per task | In main summary | |
| |---|---|---:|---:|---:|---:| |
| | Cosmos3-Nano | with_negative | 9 | 900 | 100 | 90 | |
| | Cosmos3-Nano | without_negative | 9 | 90 | 10 | 90 | |
| | Cosmos3-Super-I2V | with_negative | 9 | 900 | 100 | 90 | |
| | Cosmos3-Super-I2V | without_negative | 9 | 90 | 10 | 90 | |
| | Genie 3 | without_negative | 9 | 9 | 1 | 9 | |
| | Seedance 2.0 | without_negative | 9 | 90 | 10 | 90 | |
| | Wan2.2 | with_negative | 9 | 900 | 100 | 90 | |
| | Wan2.2 | without_negative | 9 | 90 | 10 | 90 | |
| | Wan2.7 | with_negative | 9 | 90 | 10 | 90 | |
| | Wan2.7 | without_negative | 9 | 90 | 10 | 90 | |
|
|
| The paired summary cohort contains 810 seeded videos (nine model/condition groups × nine task settings × ten seeds). Together with the nine Genie 3 videos, `in_main_summary=true` selects 819 records. The remaining 2,430 records are additional seeds from the three local-model physical-negative-prompt experiments. Samples shared by the two cohorts are stored only once. The paired seed numbers are **1, 12, 23, 34, 45, 56, 67, 78, 89, 100**. |
|
|
| Video and trajectory payloads occupy 2,705,198,614 bytes (2.519 GiB), excluding indices and documentation. |
|
|
| ## Layout |
|
|
| ```text |
| world_models/v1/ |
| README.md |
| index.parquet |
| index.csv |
| index.jsonl |
| checksums.sha256 |
| validation.json |
| metadata/ |
| seed_manifest.json |
| paired10_selection.json |
| videos/<model>/<condition>/<task>/seed_NNN.mp4 |
| trajectories/<model>/<condition>/<task>/seed_NNN.json |
| ``` |
|
|
| Genie 3 uses `video_001` instead of `seed_NNN`. Paths in the index are relative to this `v1/` directory. Task IDs and materials are standardized in release paths; `original_task` preserves source identifiers, including `C5-3`, `C7-1`, and `C19`. Genie source filenames spelling `mental` are mapped to `metal` in the index and destination paths only. |
|
|
| ## Reading the release |
|
|
| ```python |
| from pathlib import Path |
| import json |
| import pandas as pd |
| |
| root = Path('world_models/v1') |
| index = pd.read_parquet(root / 'index.parquet') |
| main = index[index['in_main_summary']] |
| record = main.iloc[0] |
| video = root / record['video_path'] |
| trajectory = json.loads((root / record['trajectory_path']).read_text()) |
| ``` |
|
|
| `index.jsonl` provides the same records without a Parquet dependency. Original random seeds are strings, preserving positive 64-bit values exactly. `seed_number` is a one-based position in the original 100-seed manifest; `seed_index` is zero-based. For the API models, `generation_seed`/`api_seed` record the submitted seed (original seed modulo 2**31), while `original_seed` records the shared pool value. Sharing a pool position does not imply identical random streams across models. For Genie 3, seed fields are null; `video_001` is an inference label, not a known seed. |
| |
| ## Full trajectory semantics |
| |
| Trajectory JSON files are copied without reformatting or changing their values. `mid_point_traj.x`/`.y` contain mask-centroid pixel coordinates keyed by original zero-based video frame numbers. `box_center_traj` contains bounding-box centers. Pixel origin is top-left; x increases rightward and y downward. Native video resolution, frame rate, and frame count are retained. No video resizing, re-encoding, frame resampling, trajectory cropping, smoothing, interpolation, or metric filtering is performed for this release. |
| |
| “Full” means the original extraction result over the entire video, not a guarantee of detections in every frame. Missing/empty masks leave absent coordinate keys. The original `frame_count` is the total video frame count; `tracked_frames`, `missing_frames`, and `tracking_coverage` in the index describe observation coverage. Low-coverage and physically unmeasurable videos are retained. Tracking coverage is **not** generation validity, metric eligibility, or physical fidelity; those require separate task-specific analysis. Full-video JSON is included even when a physical metric cannot be computed. |
|
|
| Original JSON may retain historical absolute `source_video` paths and references to local visualization/annotation files. These fields are provenance, not portable release paths. Use the index's `video_path` and `trajectory_path` to resolve released files. Visualization images, cropped analysis trajectories, apex-only event files, metrics, real-world recordings, checkpoints, and historical 5/25-seed/debug outputs are outside this package. |
|
|
| ## Verification and provenance |
|
|
| Each payload was copied into a new regular file, not a symlink or hardlink. Source size, modification time, device, and inode were checked before and after copying. Source-read SHA256 was independently compared with the destination SHA256; recorded API/Genie video hashes were also checked when available. The MP4 stream dimensions, native FPS, and container frame count were checked against its trajectory metadata using ffprobe. Coordinate keys, finite values, frame ranges, model/task identity, expected seed sets, and seed-manifest identity were checked. This build checks video stream metadata; it does not claim a new full-frame decode or a new manual target-identity audit. |
|
|
| `checksums.sha256` covers all published files except the checksum list itself. From this directory, run `sha256sum -c checksums.sha256` to verify the release. `validation.json` records counts and checks. Local source-to-release mapping, source document snapshots, and the builder remain in the sibling `_local/` directory outside the publishable `world_models/` tree. Upload only `world_models/` when adding this package to the existing dataset repository. |
|
|
| Authoritative local scope references: `WM_SAM3_100seeds.md` and `每个任务轨迹处理的规则.md`, with the current Genie 3 input manifest. This package selects the latest documented raw extraction roots and does not select task-cropped analysis outputs. |
|
|
