| --- |
| pretty_name: ProPriMem-State |
| tags: |
| - robotics |
| - memory |
| - state |
| - order |
| - recovery |
| --- |
| # ProPriMem-State |
|
|
| 21 tasks · 1,007 trajectories · 2,703 videos · 618 stage-annotation files. |
|
|
| | ID | Task | Trajectories | Source / domain | |
| |---|---|---:|---| |
| | S01 | [Cover and uncover blocks](State/01-cover_uncover_blocks/) | 50 | MINT-SJTU/ProPriMem · simulation | |
| | S02 | [Put a block back](State/02-put_back_block/) | 50 | MINT-SJTU/ProPriMem · simulation | |
| | S03 | [Move blocks to a green mat and back](State/03-place_block_mat/) | 50 | MINT-SJTU/ProPriMem · simulation | |
| | S04 | [Try battery orientations](State/04-battery_orientation_trials/) | 50 | MINT-SJTU/ProPriMem · simulation | |
| | S05 | [Try block arrangements](State/05-block_arrangement_trials/) | 50 | MINT-SJTU/ProPriMem · simulation | |
| | S06 | [Rearrange blocks](State/06-rearrange_blocks/) | 50 | MINT-SJTU/ProPriMem · simulation | |
| | S07 | [Pick and place a bowl once](State/07-pick_place_bowl_once/) | 50 | MINT-SJTU/ProPriMem · simulation | |
| | S08 | [Pick and place a bottle once](State/08-pick_place_bottle_once/) | 98 | MINT-SJTU/ProPriMem · simulation | |
| | S09 | [Swap two bowls](State/09-swap_two_bowls/) | 88 | MINT-SJTU/ProPriMem · simulation | |
| | S10 | [Rotate three bowls](State/10-rotate_three_bowls/) | 82 | MINT-SJTU/ProPriMem · simulation | |
| | S11 | [Put three coins into a piggy bank](State/11-put_three_coins_into_piggy_bank/) | 32 | EvoMind-PiperX · real robot | |
| | S12 | [Inspect four objects and return them](State/12-inspect_four_objects_and_return/) | 35 | EvoMind-PiperX · real robot | |
| | S13 | [Inspect three drawers](State/13-inspect_three_drawers/) | 35 | EvoMind-PiperX · real robot | |
| | S14 | [Weigh four objects and return them](State/14-weigh_four_objects_and_return/) | 35 | EvoMind-PiperX · real robot | |
| | S15 | [Inspect three test tubes](State/15-inspect_three_test_tubes/) | 49 | EvoMind-PiperX · real robot | |
| | S16 | [Press five buttons in any order](State/16-press_five_buttons_in_any_order/) | 35 | EvoMind-PiperX · real robot | |
| | S17 | [Pour water into three cups](State/17-pour_water_into_three_cups/) | 33 | EvoMind-PiperX · real robot | |
| | S18 | [Inspect five upside-down cups](State/18-inspect_five_upside_down_cups/) | 34 | EvoMind-PiperX · real robot | |
| | S19 | [Shake three sealed bottles](State/19-shake_three_sealed_bottles/) | 35 | EvoMind-PiperX · real robot | |
| | O15 | [Flip two cups from left to right with recovery](Order/15-flip_two_cups_left_to_right_with_recovery/) | 35 | EvoMind-PiperX · real robot · recovery | |
| | S20 | [Pour water into three cups with repeated pouring](State/20-pour_water_into_three_cups_with_repeated_pouring/) | 31 | EvoMind-PiperX · real robot · repeated-pouring anomaly | |
|
|
| ## Format and splits |
|
|
| Each task under `State/` or `Order/` retains its original `metadata.json`, `metadata.jsonl`, videos and, |
| when present, stage annotations. Paths in each trajectory index are task-relative. Resolution, |
| frame rate and available camera views are preserved per trajectory; they are not uniform |
| across all source datasets. All views of a trajectory share its frozen split. |
|
|
| Dataset v1.4.1 · Split v1.4.1 · **803 train / 204 val**. Explicit sample lists in |
| [manifest.json](manifest.json) are authoritative. All surviving assignments are preserved. |
|
|
| ## Sources and annotation status |
|
|
| S01–S10 are copied from [MINT-SJTU/ProPriMem](https://huggingface.co/datasets/MINT-SJTU/ProPriMem/tree/e04d7f2261cbff363b8ec3809a60cd5e6f2c7df7/State): |
| 618 simulation trajectories with their original stage annotations. S01–S06 originate from |
| [TianxingChen/RMBench](https://huggingface.co/datasets/TianxingChen/RMBench); S07–S10 originate |
| from [htrbao/LIBERO-MEM_lerobot](https://huggingface.co/datasets/htrbao/LIBERO-MEM_lerobot). |
| Original source links, instructions, hashes and domain labels remain in each trajectory index. |
| The MINT repository is unchanged by this copy. |
|
|
| S11–S19 contain 323 real-robot PiperX trajectories collected on 2026-10-31, 2026-11-02 and |
| 2026-11-03, with 969 videos at 1280×720 and 30 FPS. These recordings are **unannotated**: |
| `has_annotation=false`; stage boundaries and task-success labels are not inferred. The environment |
| camera is `main_view`, with wrist cameras `left_view` and `right_view`. Task names have no State |
| prefix. Raw robot state/action Parquet files are not included. |
|
|
| ## Selection and validation |
|
|
| User-reviewed exclusions remain unchanged: piggy-bank batch `20261031_153227` #3/#15/#16 and |
| batch `20261031_161309` #18/#19/#20; tea batch `20261103_134047` #0/#1; upside-down-cup batch |
| `20261103_142808` #0. These refer to original source episode indices. The bottle task describes |
| three bottles. The test-tube task combines valid batches of 11 + 3 + 35 trajectories. |
|
|
| The 969 PiperX videos are frame-exact H.264 stream-copy cuts, checked by full decoding against |
| source frame counts, timestamps and decoded YUV pixels. Imported State files are copied without |
| re-encoding or changing annotations; hashes and frozen split assignments are verified. |
|
|
| ## Order recovery addition |
|
|
| O15 adds 35 real-robot trajectories and 105 videos from the 2026-11-03 two-cup recovery task. |
| It belongs to **Order**, with `task_variant=recovery`. Including the S20 review exclusions below, |
| State contains 20 tasks / 972 trajectories; Order contains 1 task / 35 trajectories. |
| O15 retains its original files and 28 train / 7 val samples. |
|
|
| Flip the left cup over and back. While handling the right cup, knock over the left cup, |
| restore it, then finish flipping the right cup over and back. The user confirmed the actual |
| two-cup setup. The corrected text is stored in `source_instruction`; the original three-cup |
| text remains in `recorded_source_instruction`. Raw source recordings are unchanged. |
|
|
| All 35 reviewed trajectories are retained and unannotated. No recovery-stage boundaries or |
| success labels are inferred. The 105 new videos preserve 1280×720 at 30 FPS and exact source |
| frame boundaries, verified by full decoded-pixel, frame-count and timestamp checks. |
|
|
|
|
| ## Repeated-pouring anomaly addition — 2026-11-05 |
|
|
| S20 contains 31 real-robot trajectories (93 videos) of pouring water into three cups |
| with repeated pouring into the same cup. This is a **State anomalous variant**, not a |
| recovery task: `task_variant=repeated_pouring`, `anomaly_type=repeated_pouring_same_cup`, |
| `base_task_id=S17`. Keep it separate from normal demonstrations when selecting data. |
|
|
| The nominal instruction matches S17: "Pour water into each of three identical cups in any order." |
| The anomalous behavior is described in separate metadata fields; the original recorded task |
| text is retained in `recorded_source_instruction`. Manual review on 2026-11-06 excluded |
| source episodes `000007`, `000011`, `000015` and `000033`, including all three camera views |
| and their metadata entries. Original episode numbers are preserved. The remaining videos are |
| unannotated: stage/progress boundaries and per-episode success or failure labels have not |
| been assigned. The variant is based on the collector's description. |
|
|
| The retained S20 samples contain 25 train / 6 val with their original frozen assignments. Videos preserve |
| 1280×720 at 30 FPS and exact source frame boundaries, verified by full decoded-frame, |
| timestamp and pixel comparisons. Retained videos and their split assignments are unchanged. |
| Raw LeRobot recordings are retained on the collection server; this export contains videos |
| and ProPriMem metadata, not the raw robot action/state Parquet files. |
|
|
