sim-training / README.md
xspcz's picture
Upload complete 300-cases-per-category three-camera training set
01a2b39 verified
|
Raw History Blame Contribute Delete
1.78 kB
# LabOS vortexing: 300 cases per category, three views
Target: 2,400 scenes and 7,200 videos, at 1920×1200 and 30 FPS.
The categories are success, vortex_off, tube_empty, cap_open, tube_drop,
wrong_orientation, wrong_rack, and rack_flipped.
Use `progress.json` and `pipeline_status.json` for the current generation status.
`GENERATION_COMPLETE.json` is created only after final extension validation.
Do not infer completeness from the folder name.
Each `CATEGORY/scene_NNNNNN/` contains three synchronized videos (`left.mp4`,
`front.mp4`, `right.mp4`), matching per-video JSON metadata, `metadata.json`,
`shared_state.npz`, and three QA contact sheets. `manifest.jsonl` indexes the
currently accepted videos with relative paths. `camera_rig.json` freezes the
three camera poses used by the existing set.
IDs 000000–000099 are byte-verified independent copies of the original set.
IDs 000100–000299 are newly sampled scenes with non-overlapping seed allocations.
Each scene uses one physical simulation shared by its three camera streams.
The existing rendering, collision, randomization, and metadata implementation
is frozen; operational parallelism does not lower resolution or frame rate.
Known inherited exception: `rack_flipped/scene_000070` has a final rack tilt
of 71.61°, below the final audit's ≥80° threshold. It remains unchanged and
flagged in `inheritance.json`. New rack-flip scenes are rejected below 80°.
The final full-dataset integrity audit retains this exception; the separate
extension audit checks that no new failures or inherited-file changes occurred.
Frustum coverage is audited, but that is not an exhaustive visual-occlusion
or artifact certification.
This generation is local only. It does not resume the paused Hugging Face upload.