Datasets:

ml-resources
/

droid-atomic-actions

Name: droid-atomic-actions
Creator: resources
License: https://choosealicense.com/licenses/cc-by-4.0/

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

The dataset viewer is not available for this dataset.

Cannot get the config names for the dataset.

Error code:   ConfigNamesError
Exception:    TypeError
Message:      Value.__init__() missing 1 required positional argument: 'dtype'
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
                  config_names = get_dataset_config_names(
                      path=dataset,
                      token=hf_token,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
                  dataset_module = dataset_module_factory(
                      path,
                  ...<4 lines>...
                      **download_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1182, in dataset_module_factory
                  raise e1 from None
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1157, in dataset_module_factory
                  ).get_module()
                    ~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 692, in get_module
                  config_name: DatasetInfo.from_dict(dataset_info_dict)
                               ~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/info.py", line 284, in from_dict
                  return cls(**{k: v for k, v in dataset_info_dict.items() if k in field_names})
                File "<string>", line 20, in __init__
                File "/usr/local/lib/python3.14/site-packages/datasets/info.py", line 170, in __post_init__
                  self.features = Features.from_dict(self.features)
                                  ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/features/features.py", line 2023, in from_dict
                  obj = generate_from_dict(dic)
                File "/usr/local/lib/python3.14/site-packages/datasets/features/features.py", line 1602, in generate_from_dict
                  return {key: generate_from_dict(value) for key, value in obj.items()}
                               ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/features/features.py", line 1621, in generate_from_dict
                  return class_type(**{k: v for k, v in obj.items() if k in field_names})
              TypeError: Value.__init__() missing 1 required positional argument: 'dtype'

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Mirror of GemingOne/droid-atomic-actions from ModelScope. All credit to the original authors. Not an official release from them. License unchanged (cc-by-4.0).

DROID Atomic Actions — 原子动作标注（动词+名词，588,151 段）

对 DROID（r2d2_faceblur 源）的全量细粒度原子动作标注：把 94,030 条机器人操作轨迹按夹爪开合事件 + 末端执行器减速边界切分为 588,151 个原子段，每段标注一个 动词 + 名词 标签（封闭 27 词表），如 pick cup、open drawer、approach knob。

标注体系参照 Qwen-RobotManip（arXiv 2606.17846）的原子动作 taxonomy 并做了两项扩展：移动类拆成 approach/pick/move/place 四相；引入邻段上下文帧消歧 approach 的宾语。

目录结构

annotations/
  labels.jsonl        # 588,151 条标注（每段一条）
  segments.jsonl      # 段定义：时间区间、切分信号、上下文元数据
meta/
  tasks.jsonl         # task_index -> 任务语言描述（31,306 条）
source/               # （可选镜像）LeRobot v2.1 格式源数据集
  vla-lerobot-v2.1-94030-fps10/{meta,data,videos}

对应关系（labels ↔ 源数据）

seg_id 形如 droid/ep{E:06d}#s{K:02d}，其中 E = LeRobot episode_index：

要找	位置
低维数据（state/action）	`source/.../data/chunk-{E//1000:03d}/episode_{E:06d}.parquet`
三路相机视频	`source/.../videos/chunk-{E//1000:03d}/{video_key}/episode_{E:06d}.mp4`
段的帧范围	segments.jsonl 的 `[step_start, step_end)`（10fps；parquet 第 i 行 ≙ 视频第 i 帧，严格 1:1）
任务语言	segments.jsonl 的 `task_index` → `meta/tasks.jsonl`

三个 video_key：observation.images.observation.images.{exterior_image_1_left, exterior_image_2_left, wrist_image_left}（320×180 @10fps）。

labels.jsonl 字段

{"seg_id": "droid/ep000004#s01",
 "label": {"verb": "pick", "object": "block", "verb_type": "pick", "confidence": "high"},
 "model": "gpt-5.6-luna", "prompt_version": "2e05fa7e", "ts": 1784460000,
 "usage": {"in": 5489, "out": 156, "cached": 3840}}

verb == verb_type，取值严格限于下方 27 词表；object 为单个小写英文名词。
confidence：high 79.9% / medium 19.9% / low 0.2%。

segments.jsonl 字段

切分规则：夹爪从开到合 = 一次抓取相位结束；从合到开 = 一次放置相位结束；在每个夹爪相位内再用 EEF 速度的"行进/减速"边界拆出 travel/terminal 子段（对应 approach|pick 与 move|place）。
frames 里的路径是构建机的本地产物，公开数据中请忽略；采样规则为段内 0/33/67/100% 四个时刻（frame_steps 给出确切帧号），可据此从源视频重建。

27 类封闭动词表

移动四相：approach pick move place；操作类：flip rotate toggle open close push pull insert press click strike；人手/装配扩展：attach detach pour wipe stir fold unfold braid unbraid；特殊：handover retract none。

分布（top）：pick 136,804 / approach 125,674 / place 116,135 / move 105,415 / retract 75,037 / open 5,991 / close 4,329 / push 3,766 / insert 3,066 / rotate 2,230 …

标注方法

标注模型：gpt-5.6-luna（多模态），每段输入段内 4 时刻 × 3 视角帧 + 前段末时刻、后段首末时刻的上下文帧（共 ≤18 帧）+ 夹爪/速度机械提示。
关键约束：approach X 仅当后续上下文中手爪真正接触 X 才成立；宾语必须是被交互物体，禁止旁观物体；无交互段标 retract/none。
输出经封闭词表校验 + 失败自动修复重试；全部 588,151 段零失败。

许可与致谢

标注（annotations/）：CC-BY-4.0。
源数据：DROID（CC-BY-4.0），引用 Khazatsky et al., DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset, 2024。本仓库中的 source/ 为其 r2d2_faceblur 变体的 LeRobot v2.1 格式导出（坐标映射至 droid_single_arm_left，重采样 10fps）。
动词体系参照：Qwen-RobotManip Technical Report, arXiv 2606.17846。

Downloads last month: 141

Total file size:

100 GB

Collection including ml-resources/droid-atomic-actions

From ModelScope

Collection

98 items • Updated about 7 hours ago