| license: other | |
| pretty_name: RoboTwin 2.0 put_object_cabinet Qwen short-subtask annotations | |
| language: | |
| - en | |
| task_categories: | |
| - robotics | |
| # RoboTwin 2.0 `put_object_cabinet` Qwen annotations | |
| This public artifact contains the accepted short textual subtask annotations for 50 real RoboTwin 2.0 expert episodes. It contains annotation metadata only; it does not contain the source HDF5 or videos. | |
| - Source model: `Qwen/Qwen2.5-VL-7B-Instruct` | |
| - Prompt: `put_object_cabinet_overlap_v6_hard_gt_state_feedback` | |
| - Annotation revision: `3af8b3026b544d4c7da4f56d85afdd7bca303b714f763d526f476bfc3ddff943` | |
| - Split: 40 train episodes / 10 held-out episodes | |
| - Raw stage frames: object reach/grasp 3261; parallel grasp/open 84; drawer reach/open 5496; joint insert/place 4675 | |
| `parallel_grasp_open` remains intact in the raw annotation. The current PI0.5 training view merges those 84 frames into the following `drawer_reach_open` segment, producing a three-stage effective target. Consumers should preserve both views when reviewing annotation quality. | |
| The authoritative schema, per-episode boundaries, arm roles, source checksums, and provenance are in `manifest.json`. `SOURCE_INVENTORY.sha256` covers the published payload files other than the inventory and transfer receipt themselves. | |