AnonymousUserCoRL2026/keep-the-flow-1x3

Keep the Flow — 1x3 Formation

Anonymous release of a bimanual humanoid manipulation dataset for double-blind conference review. Each episode shows a 43-DoF humanoid robot completing a 1×3 formation by picking the single object available on a side cart and placing it into the empty slot of a 3-slot table layout.

Variants

The same 24 task cases are recorded in 8 variants (≈ 33K files, 49 GB total):

Variant	Description
original	Original RGB videos with table grid markings visible
original_grid_removed	Same recordings with grid removed from the videos
recolor1 / recolor1_grid_removed	Per-case color swap palette #1 (with/without grid)
recolor2 / recolor2_grid_removed	Per-case color swap palette #2 (with/without grid)
recolor3 / recolor3_grid_removed	Per-case color swap palette #3 (with/without grid)

The 24 cases sweep:

• Shape: hexagonal prism (H), box (C), cylinder (B)
• Color: red (R), orange (O), yellow (Y), blue (B)
• Spacing: 15 cm, 20 cm, 25 cm
• Empty slot position: 1×1, 1×2, or 1×3

Each variant contains 24 case folders following the pattern T3v2_<spacing-id>-<row>_1x3_<spacing>cm/.

Format

LeRobot v2.1 dataset format. Each case folder contains:

<variant>/
├── T3v2 - Keep the Flow Enhanced[ (recolor*)].xlsx   # per-variant catalog: ID, formation,
│                                                     # spacing, table layout, cart contents,
│                                                     # and 3-level instructions (detailed / medium / simple)
└── T3v2_<id>_1x3_<spacing>cm/
    ├── case_metadata.json                  # cart & slots ground truth
    ├── meta/
    │   ├── tasks.jsonl                     # single task instruction (medium template)
    │   ├── episodes.jsonl                  # per-episode length & task text
    │   ├── info.json                       # LeRobot schema (43-DoF, 3 cameras, 15 fps)
    │   ├── modality.json
    │   ├── stats.json
    │   └── relative_stats.json
    ├── data/chunk-000/episode_*.parquet    # observation.state, action, indices
    └── videos/chunk-000/
        ├── observation.images.ego_view/episode_*.mp4       (640x360)
        ├── observation.images.left_wrist/episode_*.mp4     (360x640)
        └── observation.images.right_wrist/episode_*.mp4    (360x640)

Robot platform

• 43 DoF: left/right legs (6+6), waist (3), left/right arms (7+7), left/right hands (7+7)
• 3 RGB cameras at 15 fps (ego, left wrist, right wrist)
• Action and observation share the same 43-DoF joint-position schema
• Robot type label retained in info.json for reproducibility

Task instructions

Every episode in every variant uses a single, medium-detail natural-language instruction of the form:

Complete the 1x3 formation with {spacing} spacing. The cart on your right contains only one object: a {color} {shape}. Pick it up and place it into the empty {slot} slot.

The instruction is stored both in meta/tasks.jsonl (task index 0) and in each record of meta/episodes.jsonl.

Anonymity notice

This is an anonymized release for double-blind review. Authors, institutions, and prior publications are intentionally omitted and will be added in the camera-ready version.

License

CC-BY-4.0 — attribution will be specified upon de-anonymization.