ccwatson/metaworld_bin_transfer_randomized_noise005_50eps

MetaWorld Bin Transfer Randomized Noise005 50eps

This dataset contains scripted MetaWorld demonstrations for a three-bin colored-block transfer task. The scene has one gray source tray containing a red block and a blue block, plus a red destination tray and a blue destination tray. Each episode asks the Sawyer gripper to move one named block into one named destination tray, with prompt variants that either omit a height constraint (base) or explicitly ask the robot to keep the block high above the table (high) or low to the table (low).

The intent is to provide a clean LeRobot-format dataset for OpenPI / pi0.5 fine-tuning on precise spatial and semantic control: select the correct colored object, select the correct colored destination, avoid moving the distractor block into the requested destination, and condition motion height on language.

High-Level Facts

• HF repo: ccwatson/metaworld_bin_transfer_randomized_noise005_50eps
• Format: LeRobot v2.1 style parquet episodes, compatible with the OpenPI MetaWorld LeRobot path.
• Robot type: metaworld
• Episodes: 600
• Frames: 110088
• Tasks/prompts: 12
• Episode files: 600 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: 0:600
• Collection metadata schema version: 1
• Collection metadata written at: 2026-05-27T21:31:54+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 --output-root data/metaworld_bin_transfer_randomized_noise005_50eps --dataset-repo-id ccwatson/metaworld_bin_transfer_randomized_noise005_50eps --hub-repo-id ccwatson/metaworld_bin_transfer_randomized_noise005_50eps --block-start-xy-noise-magnitude-m 0.008 --target-xy-noise-magnitude-m 0.012 --unrecorded-action-noise-magnitude 0.05 --overwrite --publish-to-hub

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 --output-root data/metaworld_bin_transfer_randomized_noise005_50eps --dataset-repo-id ccwatson/metaworld_bin_transfer_randomized_noise005_50eps --hub-repo-id ccwatson/metaworld_bin_transfer_randomized_noise005_50eps --block-start-xy-noise-magnitude-m 0.008 --target-xy-noise-magnitude-m 0.012 --unrecorded-action-noise-magnitude 0.05 --overwrite --publish-to-hub

Runtime environment recorded by the generator:

{
  "MUJOCO_GL": "egl"
}

The original built-in dataset.push_to_hub() upload path stalled in Hugging Face's large-folder/Xet pre-upload phase. The local dataset remained intact, and the final successful upload used:

HF_HUB_DISABLE_XET=1 UV_CACHE_DIR=/tmp/openpi-uv-cache uv run python -c "from huggingface_hub import HfApi; HfApi().upload_large_folder(repo_id='ccwatson/metaworld_bin_transfer_randomized_noise005_50eps', repo_type='dataset', folder_path='data/metaworld_bin_transfer_randomized_noise005_50eps', private=False, num_workers=4, print_report=True, print_report_every=30)"

Code Provenance

Generation provenance is also 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: 2802121599c6769635285c94e698dff10d8928c6 (2802121)
• Commit subject: Basic functionality for bin transfer experiment
• Generation worktree dirty: True
• Dirty file count recorded by generator: 11
• Important generation files recorded as dirty or involved in this dataset path:
  • examples/metaworld/bin_transfer.py
  • examples/metaworld/generate_bin_transfer_dataset.py
  • examples/metaworld/eval_bin_transfer_benchmark.py
  • tests/test_bin_transfer_suite.py

Dirty files recorded by the generator:

xamples/metaworld/bin_transfer.py
examples/metaworld/eval_bin_transfer_benchmark.py
examples/metaworld/generate_bin_transfer_dataset.py
tests/test_bin_transfer_suite.py
third_party/libero
#README.md#
delete_this_serve_policy.bash
delete_this_serve_policy.bash~
logs/evals/
logs/finetunes/
outputs/

Task Design

The task grid covers all object-color and destination-color pairs, including matched and mismatched colors:

• red block to red bin
• red block to blue bin
• blue block to red bin
• blue block to blue bin

Each pair has three language/motion variants:

• base: for example, move the blue block to the blue bin
• high: for example, move the blue block to the blue bin while keeping it high above the table
• low: for example, move the blue block to the blue bin while keeping it low to the table

Task keys encode target color, destination color, and height mode, for example blue-to-blue/base-00, blue-to-blue/high-00, and blue-to-blue/low-00.

Alternate Programmatic Prompts

The parquet episodes store natural-language prompts in the LeRobot task field. An alternate programmatic prompt style is included as an exact string-to-string remapping in meta/bin_transfer_programmatic_prompt_remapping.yaml. This matches the OpenPI prompt-remapping path used by the spatial-relation datasets: train/eval can load the natural-language dataset and replace prompts at runtime without rewriting parquet episodes.

Example remapping:

"move the blue block to the blue bin": "(:action transfer-block :parameters (blue_block blue_bin) :effect (and (in blue_block blue_bin) (not (in red_block blue_bin))))"

High/low variants add a trajectory constraint, for example:

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

Training with the programmatic prompt variant uses the usual OpenPI dataset prompt-remapping configuration. Bin-transfer benchmark eval supports the same file through --prompt-remapping.

Task Table

task_index	task_key	prompt	accepted episodes	final validation successes	flagged episodes	mean steps
0	red-to-red/base-00	move the red block to the red bin	50	50	-	188.9
1	red-to-red/high-00	move the red block to the red bin while keeping it high above the table	50	50	-	198.4
2	red-to-red/low-00	move the red block to the red bin while keeping it low to the table	50	49	82	160.5
3	red-to-blue/base-00	move the red block to the blue bin	50	50	-	190.1
4	red-to-blue/high-00	move the red block to the blue bin while keeping it high above the table	50	50	-	199.2
5	red-to-blue/low-00	move the red block to the blue bin while keeping it low to the table	50	50	-	165.0
6	blue-to-red/base-00	move the blue block to the red bin	50	50	-	189.1
7	blue-to-red/high-00	move the blue block to the red bin while keeping it high above the table	50	50	-	197.7
8	blue-to-blue/base-00	move the blue block to the blue bin	50	50	-	190.2
9	blue-to-blue/high-00	move the blue block to the blue bin while keeping it high above the table	50	50	-	199.0
10	blue-to-red/low-00	move the blue block to the red bin while keeping it low to the table	50	47	145, 168, 332	160.3
11	blue-to-blue/low-00	move the blue block to the blue bin while keeping it low to the table	50	50	-	163.4

Collection Configuration

The generator selected all 12 benchmark tasks and saved 50 accepted trajectories per task.

{
  "episodes_per_task": 50,
  "max_attempts_per_task": 75,
  "max_steps": 420,
  "post_success_steps": 8,
  "seed": 42,
  "task_count": 12,
  "tasks": [
    {
      "task_key": "red-to-red/base-00",
      "target_block_color": "red",
      "destination_bin_color": "red",
      "height_mode": "base",
      "prompt": "move the red block to the 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/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",
      "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",
      "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-blue/base-00",
      "target_block_color": "red",
      "destination_bin_color": "blue",
      "height_mode": "base",
      "prompt": "move the red block to the blue 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-blue/high-00",
      "target_block_color": "red",
      "destination_bin_color": "blue",
      "height_mode": "high",
      "prompt": "move the red block to the blue bin while keeping it high above the table",
      "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-blue/low-00",
      "target_block_color": "red",
      "destination_bin_color": "blue",
      "height_mode": "low",
      "prompt": "move the red block to the blue bin while keeping it low to the table",
      "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": "blue-to-red/base-00",
      "target_block_color": "blue",
      "destination_bin_color": "red",
      "height_mode": "base",
      "prompt": "move the blue block to the 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": "blue-to-red/high-00",
      "target_block_color": "blue",
      "destination_bin_color": "red",
      "height_mode": "high",
      "prompt": "move the blue block to the red bin while keeping it high above the table",
      "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": "blue-to-red/low-00",
      "target_block_color": "blue",
      "destination_bin_color": "red",
      "height_mode": "low",
      "prompt": "move the blue block to the red bin while keeping it low to the table",
      "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": "blue-to-blue/base-00",
      "target_block_color": "blue",
      "destination_bin_color": "blue",
      "height_mode": "base",
      "prompt": "move the blue block to the blue 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": "blue-to-blue/high-00",
      "target_block_color": "blue",
      "destination_bin_color": "blue",
      "height_mode": "high",
      "prompt": "move the blue block to the blue bin while keeping it high above the table",
      "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": "blue-to-blue/low-00",
      "target_block_color": "blue",
      "destination_bin_color": "blue",
      "height_mode": "low",
      "prompt": "move the blue block to the blue bin while keeping it low to the table",
      "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
    }
  ]
}

Important configured knobs:

• episodes_per_task: 50
• max_attempts_per_task: 75
• max_steps: 420
• post_success_steps: 8
• seed: 42
• block_start_xy_noise_magnitude_m: 0.008
• target_xy_noise_magnitude_m: 0.012
• unrecorded_action_noise_magnitude: 0.05
• low_max_edge_clearance_m: 0.06
• high_min_edge_clearance_m: 0.12

Randomization

Each reset used slight Gaussian randomization:

• block initial XY positions: Gaussian magnitude 0.008 m, rejected/clipped to keep both blocks in the gray source tray and avoid overlap.
• target/destination goal XY position: Gaussian magnitude 0.012 m, rejected/clipped to stay within the requested destination tray.

Aggregate randomized-position statistics are stored in meta/bin_transfer_dataset_summary.json under reset_position_statistics.

Unrecorded Action Noise

This dataset uses the standard data-collection trick where the policy action recorded in actions is the clean scripted-policy action, but the simulator is stepped with a noised version of that action.

• Config name: unrecorded_action_noise_magnitude
• Value used here: 0.05
• Noise distribution: zero-mean Gaussian
• Action dimensions noised: indices [0, 1, 2] (XYZ motion)
• Gripper action: not noised
• Recorded action: clean policy action
• Simulator action: clean action plus Gaussian XYZ noise, clipped by the environment action bounds

The goal is to make demonstrations slightly more robust and varied without training on the noise itself.

Height Scoring

Height metrics use cube edge/vertex clearance from the table, not centroid height.

• Low mode requires target block edge clearance from the table to stay <= 0.06 m.
• High mode requires target block edge clearance to be >= 0.12 m, but only after the target block has left the source tray XY footprint. The block starts in the tray, so high-mode height is intentionally not scored while it is still over the source tray.
• Base mode requires goal success but does not enforce a height trajectory constraint.

Mode aggregate summary:

height mode	accepted episodes	final validation successes	steps	max edge clearance m	min active edge clearance m	height violations
base	200	200	min 180.0, mean 189.6, median 190.0, max 198.0	min 0.107, mean 0.113, median 0.113, max 0.119	n/a	0
high	200	200	min 192.0, mean 198.6, median 199.0, max 209.0	min 0.157, mean 0.163, median 0.163, max 0.167	min 0.144, mean 0.153, median 0.153, max 0.165	0
low	200	196	min 147.0, mean 162.3, median 162.0, max 180.0	min 0.020, mean 0.043, median 0.044, max 0.060	n/a	0

Color-pair aggregate summary:

color pair	accepted episodes	final validation successes	steps
blue-to-blue	150	150	min 147.0, mean 184.2, median 190.0, max 205.0
blue-to-red	150	147	min 147.0, mean 182.4, median 189.0, max 206.0
red-to-blue	150	150	min 153.0, mean 184.8, median 190.0, max 209.0
red-to-red	150	149	min 150.0, mean 182.6, median 189.0, max 206.0

Scene Geometry

The environment uses shallow/open tray geometry, rather than tall stock MetaWorld bins, so low trajectories are physically feasible.

Source tray geometry from episode metadata:

{
  "center_xy": [
    0.0,
    0.58
  ],
  "destinations": {
    "blue": {
      "center_xy": [
        0.17,
        0.82
      ],
      "floor_thickness_m": 0.002,
      "inner_half_extents_xy": [
        0.085,
        0.07
      ],
      "name": "destination_blue",
      "open_negative_y": true,
      "rim_height_m": 0.018,
      "rim_thickness_m": 0.012
    },
    "red": {
      "center_xy": [
        -0.17,
        0.82
      ],
      "floor_thickness_m": 0.002,
      "inner_half_extents_xy": [
        0.085,
        0.07
      ],
      "name": "destination_red",
      "open_negative_y": true,
      "rim_height_m": 0.018,
      "rim_thickness_m": 0.012
    }
  },
  "floor_thickness_m": 0.002,
  "inner_half_extents_xy": [
    0.13,
    0.085
  ],
  "name": "source_gray",
  "open_negative_y": true,
  "rim_height_m": 0.018,
  "rim_thickness_m": 0.012
}

The red and blue destination trays are shallow trays with low rims and an open negative-Y approach gap. The gray source tray also has an open negative-Y approach gap so the low policy can exit through the opening instead of lifting over a wall.

Scripted Policy

The scripted policy is collision-aware and task-conditioned. It follows this rough state machine:

1. Move above the selected block.
2. Line up in XY above the block before descending.
3. Descend straight down to grasp.
4. Close the gripper and lift/retreat from the source tray.
5. Route to the requested destination tray.
6. Lower/place according to the height mode.
7. Release and retreat.

The above-then-descend approach was added because direct approach motions could bump the block before grasping. High mode carries the block elevated once outside the source tray footprint. Low mode uses low carry and routes through the tray openings.

Quality Notes

The generator accepted 600 episodes because each reached the task success condition during rollout. The final saved summary validation passes for 596 of 600 episodes. Four low-mode episodes reached success but no longer passed final-state validation at the saved final summary, likely because the target drifted or settled outside the destination after the success moment and post-success recording window.

Flagged episode indices:

82, 145, 168, 332

These are listed in machine-readable form in meta/bin_transfer_episode_quality_flags.jsonl and in the full per-episode summaries under metadata/episode_summaries/. Consumers that require strict final-state success should filter those episode indices out.

Overall aggregate statistics:

• Accepted episodes: 600
• Final-validation-success episodes: 596
• Flagged episodes: 4
• Episode steps: min 147.0, mean 183.5, median 190.0, max 209.0
• First success step: min 138.0, mean 174.5, median 181.0, max 200.0
• Max edge clearance: min 0.020, mean 0.106, median 0.113, max 0.167 m
• Min active high-mode edge clearance: min 0.144, mean 0.153, median 0.153, max 0.165 m

File Organization

README.md
meta/info.json
meta/tasks.jsonl
meta/episodes.jsonl
meta/episodes_stats.jsonl
meta/bin_transfer_collection_metadata.json
meta/bin_transfer_dataset_summary.json
meta/bin_transfer_task_summary.json
meta/bin_transfer_episode_summaries_compact.jsonl
meta/bin_transfer_episode_quality_flags.jsonl
meta/bin_transfer_programmatic_prompt_remapping.yaml
meta/bin_transfer_metadata_manifest.json
meta/bin_transfer_upload_metadata.json
metadata/episode_summaries/episode_000.json ... episode_599.json
data/chunk-000/episode_000000.parquet ... episode_000599.parquet

The main training data lives in data/chunk-000/*.parquet. The meta/ files are intended to make the dataset self-describing. The metadata/episode_summaries/ directory contains the full rollout summary JSON for every accepted episode, including randomized positions, final block positions, tray geometry, validation metrics, clearances, and task config.

LeRobot Feature Schema

{
  "observation.state": {
    "dtype": "float32",
    "shape": [
      4
    ],
    "names": [
      "observation_state"
    ]
  },
  "observation.environment_state": {
    "dtype": "float32",
    "shape": [
      39
    ],
    "names": [
      "environment_state"
    ]
  },
  "corner4.image": {
    "dtype": "image",
    "shape": [
      224,
      224,
      3
    ],
    "names": [
      "height",
      "width",
      "channel"
    ]
  },
  "gripperPOV.image": {
    "dtype": "image",
    "shape": [
      224,
      224,
      3
    ],
    "names": [
      "height",
      "width",
      "channel"
    ]
  },
  "actions": {
    "dtype": "float32",
    "shape": [
      4
    ],
    "names": [
      "actions"
    ]
  },
  "timestamp": {
    "dtype": "float32",
    "shape": [
      1
    ],
    "names": null
  },
  "frame_index": {
    "dtype": "int64",
    "shape": [
      1
    ],
    "names": null
  },
  "episode_index": {
    "dtype": "int64",
    "shape": [
      1
    ],
    "names": null
  },
  "index": {
    "dtype": "int64",
    "shape": [
      1
    ],
    "names": null
  },
  "task_index": {
    "dtype": "int64",
    "shape": [
      1
    ],
    "names": null
  }
}

Notes:

• corner4.image: external camera RGB observation.
• gripperPOV.image: wrist/gripper camera RGB observation.
• observation.state: first four MetaWorld robot state values.
• observation.environment_state: full 39-D MetaWorld observation used by this fork's MetaWorld LeRobot path.
• actions: clean scripted policy action, not the noised simulator-step action.
• task_index: integer index into meta/tasks.jsonl.

Intended Use

This dataset is intended for OpenPI MetaWorld fine-tuning, especially pi0.5 experiments that need precise language-conditioned spatial behavior. It is a new bin-transfer suite, separate from the older cardinal/diagonal spatial-relation generator. It should be useful for training or evaluating whether a policy can preserve object identity, destination identity, distractor handling, and high/low path constraints under slight reset and actuation perturbations.

Not Intended For

This is not a general-purpose physical robot dataset. It is synthetic MetaWorld/MuJoCo data from a scripted controller. The action labels are simulator actions, the objects are simple colored blocks, and the scene geometry is intentionally engineered for shallow-tray high/low feasibility.

Suggested Filtering

For strict final-state training or eval, filter out the flagged episodes listed in meta/bin_transfer_episode_quality_flags.jsonl:

82, 145, 168, 332

For broader behavior cloning, using all 600 episodes is reasonable if transient success plus realistic post-place settling is acceptable.

Reproducibility Checklist

• Use commit 2802121599c6769635285c94e698dff10d8928c6 plus the dirty-file information in meta/bin_transfer_collection_metadata.json.
• Use MUJOCO_GL=egl for headless rendering.
• Use the exact generation command above.
• Use the collection metadata sidecar to recover task specs, thresholds, randomization magnitudes, and action-noise semantics.
• Use meta/bin_transfer_dataset_summary.json for aggregate validation and task-balance checks.
• Use metadata/episode_summaries/*.json for per-episode debugging.