Datasets:

---

GT-111

/

robotwin_v3_eef

Like 0

Tidy of GT-111/robotwin_v3_eef per results/hub-alignment-plan.md.

All children ship a norm_stats.json whose state and action statistics are 20 wide while the columns are 16 wide (20 = two arms × xyz + rot6d + gripper, a layout this release does not store); the declared state mean's quaternion component is 0.025 where the data's is 0.710. This regenerates it from each child's own observation.state and action columns (6077247 state rows in this run), with mask: transition_action_valid naming the rows the action statistics cover, and not the versioned schema value another release ships. It also states the two time axes in the card, records fps_semantics: "storage" and invalid_transitions: "masked" in meta/info.json, drops eef_schema and xq_action_alignment (no action_target key: meta/action_contract.json and the card carry that statement), and separates task from collection variant in the child directories (<task>-demo_clean_collect_200-50 → <task>__demo_clean_collect_200-50), keeping the clean/aug splits. The variant suffix is not removable: clean holds both demo_clean_collect_200-50 and piper_clean_50-50, and it names the recording robot. No Parquet and no video is re-uploaded; the moves are server-side LFS copies.

Renamed keys

• collection_manifest.json and dataset_receipt.json: schema: "xq_collection_manifest_v2" / "xq_dataset_receipt_v2" -> schema: "collection_manifest" / "dataset_receipt".

• meta/action_contract.json's schema_id: "eef_absolute_next_observation_wxyz_v2" is dropped, and so is the same value in collection_manifest.json's and dataset_receipt.json's top-level action_contract field; what it encoded is carried by action_units, by each child's meta/action_contract.json (whose path the manifest lists as action_contract_path) and by the card sentence.

• codebase_version: "v3.0" is unchanged: that field names LeRobot's own format version, which the naming rule exempts.