rocochallenge2025/roco2-taskboard-assembly

Task Board Assembly — R1 Lite and R1 Pro

This dataset contains 130 simulated assembly episodes: 108 fully successful and 22 partially successful. Each episode attempts the same nine-part task board assembly using a dual-arm scripted controller. Each recorded step contains three RGB views, measured robot state, and the issued action at 10 Hz.

The board is centered and fixed within each robot configuration. The head camera uses the robot's native mounting location. Variation is limited to a small XY translation of one pickup part per episode. These are demonstrations from a fixed workcell with local variation; they do not measure broad scene or real-world generalization.

Available datasets

Directory	Episodes	Frames	State dimensions	Train / validation / test
Lite/full_success	58	79,134	40	48 / 8 / 2
Lite/partial_success	9	12,637	40	7 / 1 / 1
Pro/full_success	50	81,069	44	42 / 2 / 6
Pro/partial_success	13	21,537	44	11 / 0 / 2

Total: 194,377 synchronized steps, approximately 5 hours 24 minutes at 10 Hz. Each step has all three camera views; the step count is not multiplied by the number of cameras.

Repository layout

README.md
dataset_summary.json
Lite/
  README.md
  full_success/{data/,meta/,videos/,splits.json,episode_labels.jsonl}
  partial_success/{data/,meta/,videos/,splits.json,episode_labels.jsonl}
Pro/
  README.md
  full_success/{data/,meta/,videos/,splits.json,episode_labels.jsonl}
  partial_success/{data/,meta/,videos/,splits.json,episode_labels.jsonl}

Each leaf is a complete, independent LeRobot v3.0 dataset. Episode indices start at zero separately in each leaf. The episode_id label combines robot, category and index for a globally unique identifier. data/, meta/ and videos/ are byte-for-byte copies of the validated source datasets, preserving all episode boundaries and video offsets. Videos use H.264 MP4; numeric data and episode metadata use Parquet.

Observations and actions

• RGB cameras: observation.images.head, observation.images.left_hand, observation.images.right_hand; 320 x 240 pixels, 10 Hz. LeRobot returns float RGB tensors with shape (3, 240, 320).
• State: observation.state, float32, 40 dimensions for Lite and 44 for Pro. It contains left and right measured end-effector positions and wxyz quaternions, left and right arm joint positions, left and right arm joint velocities, and two normalized gripper positions. Lite has six arm joints per side; Pro has seven.
• Action: action, float32, 14 dimensions: left [x,y,z,rx,ry,rz,gripper], then right in the same order. Cartesian poses are absolute targets obtained by forward kinematics of the commanded joint positions; rotations are rotation vectors in radians, not Euler angles or quaternions. Grippers are normalized to [0,1] (0 closed, 1 open). Positions are in meters. Exact element names are in each meta/info.json.
• Timing: the measured state and RGB observation precede the issued action at the same 10 Hz control step. Actions are commands, not achieved next-step poses. Frame timestamps start at zero within each episode.
• Depth is not included. task_index maps to the per-part instruction in meta/tasks.parquet.

Assembly order: gear_60teeth, gear_20teeth, rod_16mm, bolt_8mm, usb_a, hdmi, pin, battery_size1, battery_size5.

Gears and batteries use physical release with seating checks. Rod, bolt, USB-A, HDMI and pin use the task's snap-attachment mechanism. Success for these pieces therefore includes this simulation assistance.

Episode labels

Each leaf has episode_labels.jsonl, one JSON object per episode, in episode-index order. These labels are additional metadata; they are not automatically injected into samples returned by LeRobot.

• episode_index, episode_id, dataset_path, robot, quality, split: identity and grouping. dataset_path is relative to repository root.
• frames, dataset_from_index, dataset_to_index: length and half-open frame range within that leaf.
• task_n_pass, task_total: original task score out of nine.
• physical_adjusted_n_pass: raw passes that also pass the extra physical seating checks where applicable.
• failed_parts, physical_failed_parts, per_part_pass: raw task failures and separately identified physical failures.
• physical_seating: actual measurements, per-part tolerances, and results for the four precision parts.
• termination, termination_error_type, termination_error: normal completion or the recorded controller error. Error fields are null for normal completion.
• all_nine_parts_attempted, recorded_parts, part_completion_reasons: attempted parts and how each operation ended.
• pickup_variation: random seed, fixed-board flag and applied offsets in meters. Missing parts in pickup_offsets_m have zero additional offset.
• task_grading: original per-part grading measurements and missing-part count.

Full-success episodes have task score 9/9, all physical seating checks passing and normal completion. All 22 partial-success episodes in this release have raw score 8/9; no raw 7/9 episodes were collected. In Pro partial-success, two episodes have physical-adjusted score 7/9 and eleven have 8/9. Six Pro partial episodes terminate with a controller error on the final component after all nine components have recorded state/action frames. The other seven Pro partial episodes and all nine Lite partial episodes end normally. Failure labels must be considered when selecting demonstrations for imitation learning.

Splits and normalization

Read each leaf's splits.json and pass its episode indices to LeRobot. The existing assignments are preserved. Pro/partial_success has no validation episodes. meta/info.json retains LeRobot's original storage split (train: 0:N); that field does not define the train/validation/test experiment split in this release. Similarly, the generic Hugging Face Parquet loader does not automatically apply splits.json or synchronize MP4 frames.

Normalization statistics in meta/stats.json are the original collection statistics over the complete leaf, including held-out episodes. Recompute statistics from the selected training episodes if your evaluation requires training-only normalization. The split describes nearby variations of the same fixed workcell, not unseen layouts.

Load locally

From the downloaded repository directory, select one leaf and its desired split:

import json
from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset

root = Path("taskboard-assembly-r1/Lite/full_success").resolve()
splits = json.loads((root / "splits.json").read_text())
dataset = LeRobotDataset(
    repo_id="local/taskboard-assembly-r1",
    root=root,
    episodes=splits["train"],
    video_backend="pyav",
)
sample = dataset[0]
print(sample["observation.state"].shape)
print(sample["observation.images.head"].shape)

Download one leaf from Hugging Face

Replace YOUR_NAMESPACE with the actual publishing account or organization after upload:

import json
from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo_id = "YOUR_NAMESPACE/taskboard-assembly-r1"
subset = "Pro/full_success"
repository = Path(snapshot_download(
    repo_id=repo_id,
    repo_type="dataset",
    revision="main",  # pin a release commit for reproducible training
    allow_patterns=[f"{subset}/**", "README.md", "dataset_summary.json", "Pro/README.md"],
))
root = repository / subset
splits = json.loads((root / "splits.json").read_text())
dataset = LeRobotDataset(
    repo_id=repo_id,
    root=root,
    episodes=splits["train"],
    video_backend="pyav",
)

The repository root groups four datasets; it has no root-level LeRobot meta/. Download the complete selected leaf first and use the root argument as above. Use the local loading route for this nested layout rather than relying on top-level automatic download or streaming.

Tested packaging environment: lerobot 0.4.4, huggingface_hub 0.35.3, datasets 4.8.5, pyarrow 25.0.1, torch 2.7.0, av 15.1.0.

中文摘要

本数据集包含 Lite 67 条和 Pro 63 条，共 130 条。每个机器人按全成功与部分成功分开存放，各子目录均可独立用 LeRobot 读取。三路相机为 head、左手和右手，320×240、10 Hz；Lite 状态 40 维，Pro 状态 44 维，动作均为 14 维。

部分成功的原始评分均为 8/9。Pro 中有 2 条物理审核后为 7/9，6 条在最后一个零件操作时因控制器错误结束，标签中保留了这些区别。训练时请读取 splits.json 并显式选择 episode，避免将保存格式中的默认 train 范围误认为实验划分。

License

A redistribution license has not been specified for this export. This package does not assign a new license to the data or robot assets.