jzpeng/theconstruct-ai_sort_trash_real_2

Sort Trash Real 2 TsFile

This repository is an Apache TsFile conversion of theconstruct-ai/sort_trash_real_2, a LeRobot v2.1 robot dataset for the task “sort the trash”.

Modalities: Time-series. The converted artifact contains numeric robot state, action, teleoperation/planner signals, frame timing, and episode/task tags. Camera videos remain in the original Hugging Face dataset.

Source dataset

• Original dataset: theconstruct-ai/sort_trash_real_2
• Original owner/uploader: The Construct AI (theconstruct-ai), the sole contributor shown in the pinned repository history
• Source revision: ff96af085cd159ddb4d9e00bd83cb857bfe1fdd7
• Task: sort the trash
• Split: train
• Scale: 27 episodes, 47,528 frames, 1 task, 50 fps
• Source frames: 27 Parquet shards (67,171,662 bytes)
• Source metadata: LeRobot codebase v2.1; meta/info.json reports 27 episodes, 47,528 frames, and discarded_episode_indices=[0,3] (all source Parquets were retained)
• License/paper: the source repository does not declare a license, paper, or formal citation

Videos

Videos are not copied into this TsFile repository. The source has 27 frame-aligned MP4 files (94,897,303 bytes) in videos/, specifically videos/chunk-000/observation.images.ego_view/. The source template is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Numeric rows remain aligned with frames by episode_index and frame_index.

Converted artifact

• TsFile: data/sort_trash_real_2_train.tsfile (33,578,812 bytes)
• Table: sort_trash_real_2_train
• Rows: 47,528
• Devices: 27 episode/task TAG combinations
• Time precision: milliseconds
• Metadata: source meta/ is mirrored; meta/info.json documents the conversion, source revision, compression profile, and video policy

Schema

Time is an INT64 millisecond timeline synthesized as round(timestamp * 1000) and restarts at zero for each episode. The redundant source timestamp field is dropped after synthesis.

TAG columns:

• episode_index
• task_index

Scalar FIELD columns:

• frame_index
• sample_index
• teleop_delta_heading
• teleop_smpl_frame_index
• teleop_stream_mode
• teleop_planner_mode
• teleop_planner_speed
• teleop_planner_height

Flattened vector FIELD groups (single-precision FLOAT; source dots become underscores):

• observation.state -> observation_state_0 ... observation_state_42 (43 FLOAT fields)
• observation.eef_state -> observation_eef_state_0 ... observation_eef_state_13 (14 FLOAT fields)
• action.wbc -> action_wbc_0 ... action_wbc_42 (43 FLOAT fields)
• observation.root_orientation -> observation_root_orientation_0 ... observation_root_orientation_3 (4 FLOAT fields)
• observation.projected_gravity -> observation_projected_gravity_0 ... observation_projected_gravity_2 (3 FLOAT fields)
• observation.cpp_rotation_offset -> observation_cpp_rotation_offset_0 ... observation_cpp_rotation_offset_3 (4 FLOAT fields)
• observation.init_base_quat -> observation_init_base_quat_0 ... observation_init_base_quat_3 (4 FLOAT fields)
• action.motion_token -> action_motion_token_0 ... action_motion_token_63 (64 FLOAT fields)
• teleop.smpl_joints -> teleop_smpl_joints_0 ... teleop_smpl_joints_71 (72 FLOAT fields)
• teleop.smpl_pose -> teleop_smpl_pose_0 ... teleop_smpl_pose_62 (63 FLOAT fields)
• teleop.body_quat_w -> teleop_body_quat_w_0 ... teleop_body_quat_w_3 (4 FLOAT fields)
• teleop.target_body_orientation -> teleop_target_body_orientation_0 ... teleop_target_body_orientation_5 (6 FLOAT fields)
• teleop.left_hand_joints -> teleop_left_hand_joints_0 ... teleop_left_hand_joints_6 (7 FLOAT fields)
• teleop.right_hand_joints -> teleop_right_hand_joints_0 ... teleop_right_hand_joints_6 (7 FLOAT fields)
• teleop.left_wrist_joints -> teleop_left_wrist_joints_0 ... teleop_left_wrist_joints_2 (3 FLOAT fields)
• teleop.right_wrist_joints -> teleop_right_wrist_joints_0 ... teleop_right_wrist_joints_2 (3 FLOAT fields)
• teleop.planner_movement -> teleop_planner_movement_0 ... teleop_planner_movement_2 (3 FLOAT fields)
• teleop.planner_facing -> teleop_planner_facing_0 ... teleop_planner_facing_2 (3 FLOAT fields)
• teleop.vr_3pt_position -> teleop_vr_3pt_position_0 ... teleop_vr_3pt_position_8 (9 FLOAT fields)
• teleop.vr_3pt_orientation -> teleop_vr_3pt_orientation_0 ... teleop_vr_3pt_orientation_17 (18 FLOAT fields)

The source index scalar is renamed to sample_index. No numeric rows or vector dimensions were intentionally dropped.

Encoding and compression

The Python TsFile table writer uses the requested compact profile: FLOAT/DOUBLE -> GORILLA, INT32/INT64 and Time -> TS_2DIFF, BOOLEAN -> RLE when present, and LZ4 compression for physical columns. TAGs use the TsFile table/device TAG mechanism and are stored as strings while their source integer dtype is recorded in meta/info.json.

Usage

from tsfile import TsFileReader

reader = TsFileReader('data/sort_trash_real_2_train.tsfile')
columns = ['episode_index','task_index','frame_index','sample_index','observation_state_0','action_wbc_0']
with reader.query_table('sort_trash_real_2_train', columns, batch_size=4096) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())
reader.close()

Validation

Local validation passed: staged Parquet rows = TsFile metadata rows = 47,528; a 1,024-row query readback succeeded; 27 devices and 385 FIELD columns were present; Time range is 0-46,220 ms per episode timeline. The JSON and Markdown reports are kept outside this uploadable directory under the local work directory.