ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps

MetaWorld Baby Bin Transfer Red-to-Red High/Low Noise005 50eps

This is a small LeRobot-format MetaWorld dataset for a two-task "baby" version of the bin-transfer benchmark. The scene is the same three-tray setup used by the full bin-transfer suite: a gray source tray contains a red block and a blue distractor block, with red and blue destination trays on the table. This dataset only asks for the red block to be moved to the red bin, with language varying only the carriage-height constraint.

The purpose is to isolate whether a policy can distinguish high versus low transport behavior when object color and destination color are held fixed. It is intentionally much smaller and simpler than the full 12-task bin-transfer dataset.

High-Level Facts

• HF repo: ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps
• Format: LeRobot v2.1 style parquet episodes, compatible with the OpenPI MetaWorld LeRobot path.
• Robot type: metaworld
• Episodes: 100
• Frames: 18032
• Tasks/prompts: 2
• Final validation successes: 100 / 100
• Quality notes: 50 accepted low-carriage episodes record source-tray rim contact with the red block; all accepted episodes still pass placement and height validation.
• Episode files: 100 parquet files in data/chunk-000/
• Cameras: corner4.image and gripperPOV.image, both 224 x 224 RGB
• State: observation.state shape [4]; observation.environment_state shape [39]
• Action: actions shape [4]
• FPS: 24
• Dataset split: train: 0:100
• Dataset size on disk before upload: about 1.4 GB
• Collection metadata schema version: 1
• Collection metadata written at: 2026-05-28T17:36:44+00:00

Exact Generation Command

The dataset was generated with this command from the repository root:

MUJOCO_GL=egl uv run python examples/metaworld/generate_bin_transfer_dataset.py --episodes-per-task 50 --max-attempts-per-task 75 --max-steps 420 --benchmark-tasks-path examples/metaworld/bin_transfer_baby_benchmark_tasks.json --output-root data/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps --dataset-repo-id ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps --hub-repo-id ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps --block-start-xy-noise-magnitude-m 0.008 --target-xy-noise-magnitude-m 0.012 --unrecorded-action-noise-magnitude 0.05 --overwrite

The Python command recorded by the generator was:

/home/christopher/Documents/openpi-finetune/openpi-metaworld/.venv/bin/python3 examples/metaworld/generate_bin_transfer_dataset.py --episodes-per-task 50 --max-attempts-per-task 75 --max-steps 420 --benchmark-tasks-path examples/metaworld/bin_transfer_baby_benchmark_tasks.json --output-root data/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps --dataset-repo-id ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps --hub-repo-id ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps --block-start-xy-noise-magnitude-m 0.008 --target-xy-noise-magnitude-m 0.012 --unrecorded-action-noise-magnitude 0.05 --overwrite

Runtime environment recorded by the generator:

{
  "MUJOCO_GL": "egl"
}

The upload command used after writing this README and metadata bundle was:

HF_HUB_DISABLE_XET=1 UV_CACHE_DIR=/tmp/openpi-uv-cache uv run python -c "from huggingface_hub import HfApi; api=HfApi(); api.create_repo(repo_id='ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps', repo_type='dataset', private=False, exist_ok=True); api.upload_large_folder(repo_id='ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps', repo_type='dataset', folder_path='data/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps', num_workers=4, print_report=True, print_report_every=30)"

Code Provenance

Generation provenance is stored in meta/bin_transfer_collection_metadata.json.

• Git remote: git@github.com:cwatson1998/openpi-metaworld.git
• Git branch at generation: chris_spatial
• Git commit at generation: cd2e75ec3c63e6fa9c9dbecf3ba91ce36dbecf8c (cd2e75e)
• Generation worktree dirty: True
• Dirty file count recorded by generator: 14

Dirty files recorded by the generator:

xamples/metaworld/README.md
examples/metaworld/baku/train_baku_metaworld.py
examples/metaworld/bin_transfer_benchmark_suite.py
examples/metaworld/eval_bin_transfer_benchmark.py
examples/metaworld/generate_bin_transfer_dataset.py
tests/test_bin_transfer_suite.py
#README.md#
delete_this_serve_policy.bash
delete_this_serve_policy.bash~
examples/metaworld/bin_transfer_baby_benchmark_tasks.json
examples/metaworld/bin_transfer_baby_benchmark_tasks_metadata.json
logs/evals/
logs/finetunes/
outputs/

Task Design

Only two tasks are included:

task_index	task_key	prompt	accepted episodes	final validation successes	mean steps	mean max edge clearance (m)
0	red-to-red/high-00	move the red block to the red bin while keeping it high above the table	50	50	197.8	0.1620
1	red-to-red/low-00	move the red block to the red bin while keeping it low to the table	50	50	162.9	0.0426

Task keys encode the fixed object/destination pair plus height mode: red-to-red/high-00 and red-to-red/low-00.

Motion / Scoring Semantics

• The target object is always the red block.
• The requested destination is always the red bin.
• The blue block is a distractor and should not end in the red bin.
• high success requires the minimum cube-edge clearance from the table to stay at or above 0.12 m after the target leaves the source tray footprint and before it reaches the destination footprint.
• low success requires the minimum cube-edge clearance from the table to stay at or below 0.06 m.
• Height is measured from the cube edge/vertices, not the cube centroid.

Collection Configuration

{
  "episodes_per_task": 50,
  "max_attempts_per_task": 75,
  "max_steps": 420,
  "post_success_steps": 8,
  "seed": 42,
  "task_count": 2,
  "benchmark_tasks_path": "examples/metaworld/bin_transfer_baby_benchmark_tasks.json",
  "tasks": [
    {
      "task_key": "red-to-red/high-00",
      "target_block_color": "red",
      "destination_bin_color": "red",
      "height_mode": "high",
      "prompt": "move the red block to the red bin while keeping it high above the table",
      "programmatic_prompt": "(:action transfer-block :parameters (red_block red_bin) :constraint (carry-height high red_block table) :effect (and (in red_block red_bin) (not (in blue_block red_bin))))",
      "prompt_variants": {
        "natural": "move the red block to the red bin while keeping it high above the table",
        "programmatic": "(:action transfer-block :parameters (red_block red_bin) :constraint (carry-height high red_block table) :effect (and (in red_block red_bin) (not (in blue_block red_bin))))"
      },
      "red_block_init_pos": [
        -0.045,
        0.575,
        0.02
      ],
      "blue_block_init_pos": [
        0.045,
        0.575,
        0.02
      ],
      "low_max_edge_clearance_m": 0.06,
      "high_min_edge_clearance_m": 0.12
    },
    {
      "task_key": "red-to-red/low-00",
      "target_block_color": "red",
      "destination_bin_color": "red",
      "height_mode": "low",
      "prompt": "move the red block to the red bin while keeping it low to the table",
      "programmatic_prompt": "(:action transfer-block :parameters (red_block red_bin) :constraint (carry-height low red_block table) :effect (and (in red_block red_bin) (not (in blue_block red_bin))))",
      "prompt_variants": {
        "natural": "move the red block to the red bin while keeping it low to the table",
        "programmatic": "(:action transfer-block :parameters (red_block red_bin) :constraint (carry-height low red_block table) :effect (and (in red_block red_bin) (not (in blue_block red_bin))))"
      },
      "red_block_init_pos": [
        -0.045,
        0.575,
        0.02
      ],
      "blue_block_init_pos": [
        0.045,
        0.575,
        0.02
      ],
      "low_max_edge_clearance_m": 0.06,
      "high_min_edge_clearance_m": 0.12
    }
  ]
}

Randomization and Action Noise

This dataset uses the same noise condition as the full noise005 bin-transfer dataset:

{
  "randomization": {
    "block_start_xy_noise_magnitude_m": 0.008,
    "target_xy_noise_magnitude_m": 0.012,
    "distribution": "gaussian_with_rejection_to_tray_interior"
  },
  "action_noise": {
    "unrecorded_action_noise_magnitude": 0.05,
    "applied_to_action_indices": [
      0,
      1,
      2
    ],
    "recorded_action": "clean_policy_action",
    "simulator_step_action": "clean_policy_action_plus_clipped_gaussian_noise"
  }
}

Important detail: the actions field records the clean scripted-policy action. The simulator was stepped with that action plus clipped Gaussian noise on action dimensions 0, 1, and 2. The gripper action was not noised.

Prompt Variants

The parquet episodes store natural-language prompts in the LeRobot task field. A programmatic prompt remapping is included at meta/bin_transfer_programmatic_prompt_remapping.yaml so OpenPI training/eval can swap prompt styles without rewriting the dataset.

"move the red block to the red bin while keeping it high above the table": "(:action transfer-block :parameters (red_block red_bin) :constraint (carry-height high red_block table) :effect (and (in red_block red_bin) (not (in blue_block red_bin))))"
"move the red block to the red bin while keeping it low to the table": "(:action transfer-block :parameters (red_block red_bin) :constraint (carry-height low red_block table) :effect (and (in red_block red_bin) (not (in blue_block red_bin))))"

Metadata Included

Additional metadata files are included beyond the standard LeRobot meta/*.jsonl files:

• meta/bin_transfer_collection_metadata.json: full generator config and run provenance.
• meta/bin_transfer_dataset_summary.json: aggregate dataset, task, height-mode, and quality statistics.
• meta/bin_transfer_task_summary.json: per-task summary table in JSON.
• meta/bin_transfer_episode_summaries_compact.jsonl: one compact row per accepted episode.
• meta/bin_transfer_episode_quality_flags.jsonl: rows for accepted episodes with validation or quality notes. In this run it contains 50 low-carriage rows with collision_contacts; these rows still have validation_success: true.
• meta/bin_transfer_baby_benchmark_tasks.json: exact two-task benchmark spec used for generation/eval.
• meta/bin_transfer_baby_benchmark_tasks_metadata.json: metadata for the baby benchmark task spec.
• meta/bin_transfer_programmatic_prompt_remapping.yaml: natural-language to programmatic prompt mapping.
• meta/bin_transfer_metadata_manifest.json: SHA-256 manifest for dataset metadata files and README.md; the manifest file itself is intentionally excluded from the manifest.
• meta/bin_transfer_upload_metadata.json: upload command and upload-time provenance.
• metadata/episode_summaries/episode_*.json: full simulator/validation summary for every accepted episode.

Dataset Organization

README.md
meta/
  info.json
  tasks.jsonl
  episodes.jsonl
  episodes_stats.jsonl
  bin_transfer_collection_metadata.json
  bin_transfer_dataset_summary.json
  bin_transfer_task_summary.json
  bin_transfer_episode_summaries_compact.jsonl
  bin_transfer_episode_quality_flags.jsonl
  bin_transfer_baby_benchmark_tasks.json
  bin_transfer_baby_benchmark_tasks_metadata.json
  bin_transfer_programmatic_prompt_remapping.yaml
  bin_transfer_metadata_manifest.json
  bin_transfer_upload_metadata.json
metadata/
  episode_summaries/
    episode_000.json
    ...
data/
  chunk-000/
    episode_000000.parquet
    ...

Intended Use

Use this dataset for quick high-vs-low carriage experiments where the object and destination identity are fixed. It should be useful for smoke tests, prompt-style tests, BAKU/OpenPI data plumbing, and small LoRA runs before scaling to the full 12-task bin-transfer suite.

For full semantic coverage across red/blue objects and red/blue destinations, use ccwatson/metaworld_bin_transfer_randomized_noise005_50eps instead.