| --- |
| pretty_name: GraphVLA V7 training-cache transfer |
| task_categories: |
| - robotics |
| --- |
| # GraphVLA V7 training cache |
|
|
| Public transfer repository published at the project owner's request. |
| This is the existing verified V5 DA3/TCP LIBERO training cache reused for V7. |
| V7 mask-coverage-weighted visual pooling is computed during training, not stored here. |
| No trained policy checkpoints, credentials, or evaluation results are included. |
|
|
| The complete cache has 1,693 episodes and 273,465 dense control frames. |
| Frames contain RGB/wrist images, masks, DA3 geometry, robot TCP information, |
| normalized action chunks and provenance. The manifest uses relative paths. |
| Shards contain unmodified cache bytes and have SHA-256 checksums in MANIFEST.json. |
| Until MANIFEST.json has `complete: true`, this repository is still uploading. |
| Verify all shards and the original per-frame checksums before training. |
| Respect the upstream LIBERO dataset and dependency licenses; this transfer does |
| not grant additional rights or certify a benchmark-fairness claim. |
|
|
