README.md

metadata

license: other
pretty_name: WorldArena Track1 RoboTwin Aloha-AgileX Clean 1500
language:
  - en
task_categories:
  - text-to-video
  - robotics
tags:
  - WorldArena
  - RoboTwin2
  - Aloha-AgileX
  - robotics
  - action-to-video
  - ABot-PhysWorld
size_categories:
  - 1K<n<10K

WorldArena Track1 RoboTwin Aloha-AgileX Clean 1500

This dataset is a curated 1,500-episode RoboTwin2/Aloha-AgileX dual-arm gripper dataset prepared for WorldArena Track1-style ABot-PhysWorld SFT and A2V experiments.

It contains only the cleaned release artifacts. The original RoboTwin raw collection folders, HDF5 dumps, collection logs, smoke-test outputs, camera-debug grids, and training checkpoints are intentionally excluded.

Summary

• Episodes: 1500
• Embodiment: Aloha-AgileX dual-arm gripper
• Config: wa_clean_fixed, RT sample count 256 collection setting
• Video: 640x480 mp4
• First frame: 320x240 png
• Actions: joint14, normalized joint14, ee16, and joint14+ee16
• Camera: head camera, HDF5 verified for self-collected RoboTwin data
• SFT positives: 1500
• A2V positives: 1500
• Captions: short, WorldArena-style, and ABot-style dense captions

Trajectory length T:

• min: 76
• median: 191.0
• p95: 512.0
• max: 715

Task Distribution

task_family	count	percent
articulated_open_close	130	8.7%
button_press_click	150	10.0%
coverage_unknown	100	6.7%
dumping_pouring	60	4.0%
handover	50	3.3%
hanging	40	2.7%
lifting	100	6.7%
object_to_container	170	11.3%
pick_place	180	12.0%
ranking_arrangement	60	4.0%
rotation_orientation	70	4.7%
scanning_qrcode	80	5.3%
shaking	70	4.7%
stacking	120	8.0%
tool_use	120	8.0%

Directory Structure

episodes/rt_xxxxxx/
  observation.mp4
  first_frame.png
  action_joint14_raw.npy
  action_joint14_norm.npy
  action_ee16.npy
  action_joint14_ee16.npy
  camera_intrinsic.json
  camera_extrinsic.json
  camera_info.json
  meta.json
  quick_contact_sheet.jpg
  visual_sanity.json

manifests/
  episode_manifest.parquet
  episode_manifest.csv
  action_normalization_config.json
  worldarena_target_spec.yaml
  collection_job_summary.csv

sft_worldarena_style_caption_mix/metadata.jsonl
sft_pilot/train.jsonl
sft_pilot/val.jsonl
sft_pilot/fixed_eval.jsonl
a2v_worldarena_ee16_caption_mix/metadata.jsonl
captions_abot_style/

Metadata Formats

SFT metadata lines:

{"video":"episodes/rt_000000/observation.mp4","prompt":"...","episode_id":"rt_000000"}

A2V metadata lines:

{"video":"episodes/rt_000000/observation.mp4","prompt":"...","action_path":"episodes/rt_000000/action_ee16.npy","intrinsic_path":"episodes/rt_000000/camera_intrinsic.json","extrinsic_path":"episodes/rt_000000/camera_extrinsic.json","original_size":[480,640]}

Action Representation

• action_joint14_raw.npy: left arm 6 + left gripper 1 + right arm 6 + right gripper 1.
• action_ee16.npy: left xyz + left quaternion + left gripper + right xyz + right quaternion + right gripper.
• action_joint14_ee16.npy: concatenated 30D representation.
• Quaternion convention in A2V metadata: wxyz.
• EE local z offset used for action-map training/debug: 0.0.
• Action-map convention: robotwin_hdf5_z0.

Camera Convention

Camera source distribution:

{
  "hdf5_verified": 1500
}

Embodiment distribution:

{
  "aloha-agilex": 1500
}

For self-collected RoboTwin data, the camera convention is:

• camera: head_camera
• intrinsic: raw OpenCV K
• extrinsic: inverse of RoboTwin observation/head_camera/extrinsic_cv, exported as camera-to-world JSON for ABot/VACE utilities

Recommended Usage

SFT:

DATASET_BASE_PATH=/path/to/this_dataset
DATASET_METADATA_PATH=$DATASET_BASE_PATH/sft_worldarena_style_caption_mix/metadata.jsonl

A2V ee16:

DATASET_BASE_PATH=/path/to/this_dataset
DATASET_METADATA_PATH=$DATASET_BASE_PATH/a2v_worldarena_ee16_caption_mix/metadata.jsonl

The A2V metadata is configured for ee16 with quat_order=wxyz and ee_local_z_offset=0.0.

Notes and Limitations

• This is a generated RoboTwin2-style dataset intended for WorldArena Track1 experiments, not official WorldArena training data.
• The release excludes raw RoboTwin HDF5 and collection logs to keep the dataset compact.
• The dataset is focused on Aloha-AgileX dual-arm gripper manipulation and is not intended as a cross-embodiment dataset.
• Use the included manifests and camera/action metadata when training ABot-PhysWorld A2V models.