Nero Dual Arm: Memory-Guided Manipulation with Visual Cues
数据集简介 / Overview
Nero Dual Arm EM-WAM 是一个面向具身智能记忆能力研究的机器人操作数据集，任务为视觉信号引导、具有相邻移动约束的四盘遍历（Memory-Guided Four-Plate Traversal with Adjacency Constraints）。 场景包含四个分别位于正方形四角的盘子、一盏指示灯和一个红色方块。左右两条机械臂分别位于盘子阵列下方的左右两侧。红色方块的初始盘子可以是左下、左上、右下或右上任意一角，记为 H。H 表示当前 episode 的初始盘子，而非固定的空间位置；它在该 episode 内保持不变，并始终作为红灯指令的返回目标。
初始指示灯为白色。每次指示灯变绿时，夹爪只能将方块从当前盘子移动到一个相邻盘子，即沿正方形的一条边完成一次盘间转移，不能直接移动到对角盘子。指示灯变红时，夹爪可以从任意盘子将方块直接移回 H，该返回动作不受绿灯的相邻移动约束。后续绿灯动作仍须从方块实际所在盘子出发，满足相邻关系，直至遍历全部四个盘子。
该任务要求策略同时利用当前灯光、方块位置、盘子间的空间邻接关系及历史访问记录，并记住当前 episode 的初始盘子 H。红灯使方块返回 H，但保留此前的遍历进度；策略需要记住已访问盘子，并在后续合法的相邻移动中继续完成覆盖。因此，任务中的 memory 能力体现为跨越返回动作保持访问历史，并将历史信息用于后续路径选择。数据可用于研究视觉条件模仿学习、时序状态跟踪，以及带空间约束的记忆型操作策略。
English summary. This dataset captures a memory-dependent manipulation task with four plates at the corners of a square, a signal light, and a red block. The two robot arms are positioned on the lower-left and lower-right sides of the plate array. The block may start on any of the four corner plates. The starting plate is denoted by H: it is episode-specific, remains fixed throughout that episode, and is the return target for every red cue. The light is initially white. Each green cue permits one move from the current plate to an adjacent plate along a square edge; diagonal transfers are prohibited. A red cue permits a direct return to H from any plate, including the diagonally opposite plate. Subsequent green moves must again satisfy adjacency from the block's actual location. The task is to visit all four plates while remembering the episode-specific home plate and retaining visit history across returns to H, using that history to guide continued traversal.
场景布局 / Spatial layout
以下采用俯视示意，图中下方为机械臂所在一侧。本图仅展示 H 位于左下角的一个示例，H 也可以位于另外三个角。 在本图中，将左上角标为 P1、右下角标为 P2、右上角标为 P3。一般情况下，用 P1、P2 表示与当前 H 相邻的两个盘子（两者命名可互换），用 P3 表示 H 的对角盘子；当 H 改变时，这些符号与物理位置的对应关系也随之改变。符号不表示固定访问先后，H 始终是四个盘子之一。
P1（左上）────── P3（右上）
│ │
│ │
H（左下）────── P2（右下）
初始盘子
左机械臂 右机械臂
（阵列下方两侧）
| 当前盘子 | 绿灯允许的目标盘子 | 绿灯禁止直接到达的对角盘子 |
|---|---|---|
| H | P1 或 P2 | P3 |
| P1 | H 或 P3 | P2 |
| P2 | H 或 P3 | P1 |
| P3 | P1 或 P2 | H |
绿灯限制的是盘间转移的起点与终点；图中的边表示邻接关系，不指定夹爪在三维空间中的精确运动轨迹。红灯允许 P1 → H、P2 → H 或 P3 → H，其中 P3 → H 是允许的对角返回。
任务规则 / Task protocol
| 指示灯或任务状态 | 期望行为 | 需保留的记忆 |
|---|---|---|
| 初始白灯 | 方块位于本段任务选定的 H（任意一角），等待后续颜色指令 | 本段初始盘子 H 的位置与场景布局 |
| 每次变绿 | 从当前盘子移动到一个相邻盘子，每次只转移一条边，不跨对角线 | 已访问盘子与遍历进度 |
| 变红 | 从任意盘子直接返回本段任务的初始盘子 H；允许从对角盘子 P3 返回 | 记住 H 的位置，保留此前的访问记录 |
| 红灯后再次变绿 | 从 H 移动到 P1 或 P2，再通过后续相邻移动继续遍历 | 结合历史记录选择后续路径 |
| 已覆盖全部四个盘子 | 达到遍历完成条件；H 为初始已占据盘子，其余三个盘子均已到达 | H、P1、P2、P3 的访问记录 |
空间规则允许多条合法路径，本文不指定必须顺时针或逆时针遍历。下面的完成条件和路径用于说明任务，不表示全部收录轨迹均已经逐条人工判定成功，也不构成新增的逐帧灯光或盘子访问标签。
任务指令 / Task Prompt
以下任务指令将上述场景、灯光规则和记忆要求整合为一段完整说明，适用于四种初始盘子位置。英文版本可作为数据转换时的 --instruction 文本；其中 home plate 即本段任务的初始盘子 H。
English task prompt
Move the red block among the four plates arranged at the corners of a square. Remember the plate where the block starts as the home plate for the entire episode, and keep track of all visited plates. The light is initially white; wait for a color cue. Each time the light turns green, move the block once from its current plate to an adjacent plate along an edge of the square, never directly to the diagonally opposite plate. When the light turns red, return the block directly to the home plate from any plate, including the opposite corner, without forgetting the visited plates. On subsequent green cues, continue with adjacent moves until all four plates, including the home plate, have been visited.
中文对照
在正方形四角的四个盘子之间移动红色方块。记住方块开始时所在的盘子，将它作为整个 episode 的初始盘子 H，并记录已经访问过的盘子。指示灯初始为白色，等待颜色指令。每次灯变绿，只将方块从当前盘子移动到一个相邻盘子，即沿正方形的一条边完成一次盘间转移，不能直接移动到对角盘子。灯变红时，从任意盘子直接将方块移回 H，包括从对角盘子返回，同时保留已访问盘子的记录。之后遇到绿灯，继续按照相邻移动规则遍历，直到四个盘子全部被访问，包括初始盘子 H。
本指令已作为统一的整段任务标注写入全部 120 个 episode：meta/source_metadata/instructions.json 的 full-instructions、meta/tasks.parquet 的任务文本，以及 meta/episodes/chunk-000/file-000.parquet 的 tasks 字段。每个独立数据集的任务编号保持为 0。segment-instructions 仍为空列表，表示尚未添加带时间区间的分段标注。下面的路径示例展示这些规则如何应用于具体的灯光序列。
路径与任务示例 / Routes and task examples
下面两个示例沿用上图中 H 位于左下角的命名；将 H 设为其他角时，按相同邻接关系重新对应 P1、P2、P3 即可。
示例 1：连续绿灯指令下沿相邻盘子遍历。 每一行绿灯表示一次新的绿色移动指令：
白灯：H 已覆盖：H
绿灯：H → P1 已覆盖：H、P1
绿灯：P1 → P3 已覆盖：H、P1、P3
绿灯：P3 → P2 已覆盖：H、P1、P3、P2；遍历完成
另一条合法路径为 H → P2 → P3 → P1，每次转移同样满足相邻关系。绿灯下不能用 H → P3 或 P1 → P2 代替中间步骤，因为它们跨越对角线。
示例 2：红灯返回后，结合记忆继续相邻遍历。
白灯：H 已覆盖：H
绿灯：H → P1 已覆盖：H、P1
红灯：P1 → H 已覆盖：H、P1（保留）
绿灯：H → P2 已覆盖：H、P1、P2
绿灯：P2 → P3 已覆盖：H、P1、P2、P3；遍历完成
这里在返回 H 后选择 P2，是一条利用“P1 已访问”这一历史信息的合法续行示例，并不意味着 P1 在空间上变成不可达。如果方块位于 P3 时出现红灯，则可以直接执行 P3 → H；但之后再次变绿时，只能从 H 到 P1 或 P2，不能因为需要继续遍历就直接跳到 P3。
不同初始位置的合法路径示例。 以下每个箭头对应一次绿灯允许的相邻移动，四条路径均覆盖四个盘子：
| 本段初始盘子 H | 连续绿灯下的一条合法遍历路径 | 任意红灯时的返回目标 |
|---|---|---|
| 左下角 | 左下 → 左上 → 右上 → 右下 | 左下角 |
| 左上角 | 左上 → 右上 → 右下 → 左下 | 左上角 |
| 右下角 | 右下 → 左下 → 左上 → 右上 | 右下角 |
| 右上角 | 右上 → 右下 → 左下 → 左上 | 右上角 |
这些是可行路径示例，不规定唯一遍历方向，也不表示现有数据中四种起点的样本数量相同。例如，当 H 为右上角时，红灯允许从左下角直接返回右上角；随后绿灯只能从右上角移动到左上角或右下角。
这些路径示例用于说明任务逻辑，不对应指定 episode 的逐帧轨迹。
数据规模与格式 / Data and format
This repository contains 120 standalone, single-episode LeRobot v3.0 datasets in memory_guided_plate_traversal/episode_000000 through memory_guided_plate_traversal/episode_000119. Each folder has its own meta, data, and videos directories. The repository root is a collection of datasets, not one merged LeRobot dataset.
Total: 89,545 aligned frames, 360 color videos, 20 Hz. Each episode contains three 640×480 color camera streams, dual-arm joint state and end poses, PIKA poses, and gripper encoder observations.
observation.state concatenates the original left and right joint-position arrays (16 components total, source order and units retained). action is the next aligned frame's state; the last action repeats the last state. These action targets are inferred from observations, not recorded control commands.
All streams use nearest-sample alignment on a regular 20 Hz timeline within their common coverage. observation.source_time and observation.source_timestamps retain the target and selected original timestamps. H.264 videos retain original resolution and use CRF 18 / yuv420p.
数据路径 / Data paths
- 仓库主页：sgrespaw/Nero_Dual_Arm_EM-WAM
- 文件列表：Files and versions
- 首段示例：memory_guided_plate_traversal/episode_000000
- 末段示例：memory_guided_plate_traversal/episode_000119
远程数据集统一放在 memory_guided_plate_traversal/ 下，episode 目录使用六位补零编号。本地原始数据也已按此规则整理；本地转换结果现使用 episode0_v3 … episode119_v3 名称。下面的本地绝对路径仅适用于原始采集机器，其他使用者应替换为自己的下载目录。
| 内容 | 路径 |
|---|---|
| 原始采集数据（本地保留） | /mnt/nas/data_3_orbbec_v2/memory_guided_plate_traversal/episode_000000/ … episode_000119/ |
| 转换结果（本地） | /mnt/nas/data_3_orbbec_v2/episode0_v3/ … episode119_v3/ |
| 上传结果（仓库内） | memory_guided_plate_traversal/episode_000000/ … memory_guided_plate_traversal/episode_000119/ |
| 单段关节、动作及其他数值数据 | memory_guided_plate_traversal/episode_000000/data/chunk-000/file-000.parquet |
| 单段格式与特征定义 | memory_guided_plate_traversal/episode_000000/meta/info.json |
| 单段任务表 | memory_guided_plate_traversal/episode_000000/meta/tasks.parquet |
| 单段 episode 元数据 | memory_guided_plate_traversal/episode_000000/meta/episodes/chunk-000/file-000.parquet |
| 单路视频示例 | memory_guided_plate_traversal/episode_000000/videos/observation.images.orbbecCamera_cc1n162011t/chunk-000/file-000.mp4 |
| 源数据质量说明 | conversion_v3_tools/source_quality_notes.json |
另外两路视频目录分别为 observation.images.orbbecCamera_cc1n162014k 和 observation.images.orbbecCamera_cc1n16201th，路径结构相同。每个文件夹都是独立的单 episode 数据集，其内部 episode_index 为 0；根目录不能直接当作一个已合并的 120-episode LeRobot 数据集读取。
Loading one episode
After gaining access to this repository if required:
from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset
repo_id = "sgrespaw/Nero_Dual_Arm_EM-WAM"
folder = "memory_guided_plate_traversal/episode_000000"
cache = snapshot_download(repo_id, repo_type="dataset", allow_patterns=f"{folder}/**")
dataset = LeRobotDataset(
repo_id=repo_id,
root=Path(cache) / folder,
video_backend="pyav",
)
frame = dataset[0]
Validated using lerobot==0.4.3. Every folder passed data/schema checks, complete video frame-count checks, action-shift checks over all frames, and first/middle/last frame decoding with the official reader.
Source quality and curation
On 2026-10-16, original acquisition episodes 11, 29, 35, and 36 were removed from the current collection because of substantial joint-state sampling gaps that also affect next-state action targets. Their local raw recordings and converted datasets were deleted at the owner's request. The remaining 120 episodes were renumbered consecutively from 0 to 119, preserving acquisition order.
All episode numbers in earlier audit reports refer to the original acquisition numbering. New folder numbers are different after the first removal. For example, new episode 11 is original episode 12; new episode 29 is original episode 31; new episode 35 is original episode 39; new episode 36 is original episode 40. Use episode_index_mapping.csv or episode_index_mapping.json to trace every recording. Existing meta/conversion.json and meta/reader_validation.json remain historical provenance and may contain old IDs and paths. Each standalone dataset's internal episode_index stays 0.
The retained trajectories, action arrays, timestamps, task annotations, videos, and per-episode statistics are unchanged. The shared task prompt remains in all 120 retained datasets; segment instructions remain empty. Retention does not mean every segment passes the stricter 15 Hz ACT synchronization checks or has been behaviorally labeled successful. The remaining >100 ms alignment findings concern end-pose streams in original episodes 1, 2, 3, 10, 21, 22, and 32 (new episodes 1, 2, 3, 10, 20, 21, and 30). See conversion_v3_tools/source_quality_notes.json, which now includes both new and original IDs. Original episode 21 also has a shorter joint-state gap below the 100 ms alignment-error screening threshold. Inspect timestamps and the training feature set before use; missing measurements were not reconstructed.
Historical conversion tools and conversion_v3_tools/final_audit.json describe the original 124-recording conversion, not the curated collection. The index mapping is the current inventory. Only converted datasets are uploaded here. Retained source recordings remain local; Hub commit history still contains earlier versions of this repository.
Folder organization
All 120 converted LeRobot v3 datasets are grouped under memory_guided_plate_traversal/, with six-digit episode names:
memory_guided_plate_traversal/
├── episode_000000/
│ ├── meta/
│ ├── data/
│ └── videos/
├── episode_000001/
└── … episode_000119/
Each episode directory remains an independent single-episode LeRobot v3 dataset. Neither the repository root nor the task grouping directory is a merged LeRobot dataset root; load an individual episode directory as shown above. Removing _v3 from the directory name does not change the data format.
Local raw recordings are now grouped under /mnt/nas/data_3_orbbec_v2/memory_guided_plate_traversal/episode_000000/ through episode_000119/. Local converted datasets are renumbered at /mnt/nas/data_3_orbbec_v2/episode0_v3/ through episode119_v3/. The similarly named local task directory contains raw recordings, while the Hub task directory contains converted datasets. Historical paths and annotation notes inside provenance records describe the original conversion state. Provenance files are preserved as historical records; the later task-annotation update is documented above. Numeric data and videos remain unchanged.
- Downloads last month
- 85