| --- |
| pretty_name: "CTR Scan Object Sequential50" |
| task_categories: |
| - robotics |
| --- |
| |
| # Scan Object Sequential50 |
|
|
| Q4-A non-CTR Sequential training data containing 50 whole Scan Object |
| episodes: 25 physical L->R schedules and 25 physical R->L schedules. The |
| dataset keeps each frozen training scene seed exactly once. Even source slots |
| select L->R and odd source slots select R->L. |
|
|
| Both parent revisions share the same source program, calibrated |
| `centered_fovy90` cameras and scanning-tail control hash for every paired scene. |
| No frame-level trajectory splicing or scene replacement is used. |
|
|
| The two parent MP4 containers for each camera are retained byte-for-byte at |
| separate file indexes. Episode metadata exposes only the selected whole-episode |
| segments. Every selected video frame is decoded during qualification and image |
| statistics are recomputed from exact RGB histograms; no selected frame is |
| transcoded. |
|
|
| IdleMask is disabled. The output drops the parent-only |
| `observation.arm_active_mask` field and training must not recreate or consume |
| it. Training consumers must compute exact global normalization statistics over |
| valid samples from this immutable output revision. |
|
|
