更新时间:2026-11-13 GMT+08:00
分享

LeRobotV3数据清洗

数据清洗算子用于检测 LeRobot V3 数据集中低质量的 episode/帧(如 NaN/Inf 异常值、过短 episode、模糊/黑帧视频、动作突变、极端值、状态-动作不对齐等),生成详细报告,并对每个 episode 给出 KEEP/TRIM/DROP 决策。算子可选地基于决策构建新的清洗后数据集,全程不修改源数据集。

  • 核心能力:
    • 多种检测:支持 NaN/Inf、episode 长度、语言文本、视频质量(黑帧,模糊帧)、动作突变、极端值、状态-动作对齐等检测项。
    • 两种运行模式:filter(构建清洗后数据集,应用 DROP/TRIM)、mask(保留所有帧+打 curation_mask 标签)。
    • 统一决策策略:所有 error 级检测结果按 flagged 帧占比统一决策(KEEP/TRIM/DROP),无per-check 规则。
    • 配置驱动:通过 YAML 配置文件描述字段映射、检测项阈值,配置可放置在数据集目录或使用平台内置。
    • 使用自定义配置fk_consistency校验。
    • 基坐标系对齐:对每个数据集施加一次旋转校正,使末端位姿记录系的+x 对齐机器人基座正前方、+z 向上,从而跨数据集几何一致。
  • 当前支持机型: galaxea(星海图 R1)、a2d(智元 A2D)。其他机器人可通过 manifest 模式(数据集目录放置自定义 yaml)接入。
  • 输入:LeRobot V3格式数据集。
    支持以下两种方式指定配置(如图1所示):
    1. 通过挂载配置文件(可选):通过创建数据处理任务时的“挂载配置”参数,从OBS中挂载自定义YAML配置文件manifest.yaml,适用于根据业务需要,灵活切换配置的场景,指定算子运行所需的额外配置或数据文件,自定义YAML配置文件用于描述数据集字段映射(state_fields/action_fields/language_fields)和各检测项阈值(NaN/Inf、episode 长度、视频质量、动作突变、极端值、状态-动作对齐等)。
    2. 通过环境变量配置robot参数:通过环境变量robot指定平台内置机型配置(galaxea_r1或a2d),适用于星海图 R1 和智元 A2D 机器人,无需额外提供配置文件。

    配置优先级从高到低依次为:通过挂载配置文件 > 通过环境变量配置robot参数。高优先级配置会覆盖低优先级配置中的相同字段。对于非内置机型(非galaxea_r1/a2d),需使用manifest模式,在挂载目录中预置自定义YAML配置文件,定义字段映射、检测阈值、运动学参数(kinematic_poses)等。

    图1 两种配置方式
  • 输出:filter模式输出清洗后数据集;mask模式输出带curation_mask/curation_reason列的数据集副本。

参数配置

表1 参数说明

参数名

是否必填

默认值

可选值

说明

mode

否

mask

filter、mask

运行模式。

  • filter:构建清洗后数据集(应用 DROP/TRIM)
  • mask:保留所有帧+打 curation_mask 标签。

robot

否

manifest

galaxea_r1、a2d、manifest

指定机器人配置。

  • galaxea_r1/a2d:使用平台内置配置。
  • manifest:通用YAML manifest,需在挂载目录或者数据集根目录预置manifest.yaml。

输入说明

输入数据为LeRobot V3格式数据集,可选配置文件通过robot环境变量指定内置机型(galaxea_r1/a2d),或通过manifest模式在挂载目录放置自定义YAML配置文件。挂载配置文件优先级高于内置配置。

输入数据集目录结构

输入须为LeRobot V3格式数据集,同级可放置配置文件:

<input_dir>/  
├── <dataset>/                          # 数据集根目录(V3 格式)  
│   ├── data/  
│   │   └── chunk-000/  
│   │       └── file-*.parquet  
│   ├── meta/  
│   │   ├── info.json  
│   │   ├── stats.json  
│   │   ├── tasks.parquet  
│   │   └── episodes/  
│   └── videos/  
│       └── observation.images.*/  └── <dataset>.yaml                     # 配置文件(与数据集同名 .yaml,可选)

自定义YAML配置文件manifest.yaml

YAML配置文件manifest.yaml必须存放在当前CloudRobo工作空间OBS存储路径下(例如OBS/cloudrobo-test-203 )。

数据清洗配置文件描述数据集字段映射(state_fields/action_fields/language_fields)、检测项阈值(nan_inf/episode_length/language/video_quality/sudden_change/extreme_value/state_action_alignment)等。global.yaml提供全局默认值,数据集特定配置覆盖全局默认。

Galaxea R1(galaxea)
dataset:
  name: galaxea
  repo_id: ""
  subdir: ""
  robot_type: r1
  state_fields: [observation.state]      # 状态字段映射
  action_fields: [action]                # 动作字段映射
  language_fields: []                    # 语言字段(无则为空)
  task_from_task_index: true             # 任务文本从 task_index 解析

checks:
  nan_inf:                               # NaN/Inf 异常值检测
    enabled: true
    arrays: [state, action]
  language:                              # 语言文本检测(galaxea 默认关闭)
    enabled: false
    min_chars: 1
    max_chars: 1000
    visible_ratio_min: 0.8
    repeated_char_run_max: 40
  episode_length:                         # episode 长度检测
    enabled: true
    abs_min: 5
    abs_max: 5000
    low_percentile: 1
    high_percentile: 99
  video_quality:                          # 视频质量检测(galaxea 默认关闭)
    enabled: true
    black_mean_max: 15.0
    black_lowvar_ratio: 0.95
    blur_laplacian_min: 30.0
    blue_metric: relative                # 加这行,默认laplacian
    blur_relative_min: 1.25              # 比值下限
    blur_single_min: 40.0 

  sudden_change:                          # 动作突变检测
    enabled: true
    smoothing: savgol
    savgol_window: 7
    savgol_poly: 2
    moving_window: 5
    residual_threshold: 0.2
    accel_threshold: 1.0
    jerk_threshold: 1.5
    require_accel_or_jerk: true
    exclude_gripper: true
    gripper_dims: [6, 13]                # gripper 在 14 维 state/action 中的索引
  extreme_value:                          # 极端值检测(仅标注,不参与决策)
    enabled: true
    effect: annotate_only
    signal_keys: [state, action]
    lower_q: 0.01
    upper_q: 0.99
    alpha: 3.0
    gripper_dims_by_signal:
      state: [6, 13]
      action: [6, 13]
    gripper_unique_min: 2
  state_action_alignment:                 # 状态-动作对齐检测
    enabled: true
    smoothing: savgol
    savgol_window: 7
    savgol_poly: 2
    max_lag: 5
    direction_agree_min: 0.6  

输出说明

输出根据运行模式不同:filter模式输出清洗后数据集(DROP episode跳过,TRIM episode的连续保留区间成为独立episode);mask模式输出带curation_mask/curation_reason列的数据集副本(保留所有帧)。

filter/mask模式输出

<output>/  
├── report/
├── data/  
│   └── chunk-000/  
│       └── file-*.parquet             
├── meta/  
│   ├── info.json  
│   ├── stats.json  
│   ├── tasks.parquet  
│   ├── episodes/  
│   └── curation/  
│       └── source_mapping.parquet    
└── videos/      └── observation.images.*/

fk_consistency校验

配置支持运行fk_consistency。

约束条件:

  • 数据集中包含eef位姿数据。
  • 提供自定义配置文件并填写kinematic_poses相关配置。

以下为galaxea_r1支持fk_consistency校验的配置示例,主要新增kinematic_poses关于关节映射的内容。以及在advanced中开启fk_consistency的配置开关(如下加粗字体所示)。

# Galaxea R1 (bimanual, unified 14-dim state/action) curation config.
# Merged on top of global.yaml.
# Dataset: D:\downloads\galaxea_r1 (3 episodes, 10 fps, no language columns).
dataset:
  name: galaxea
  repo_id: ""
  subdir: ""
  robot_type: r1
  ...
  kinematic_poses:
    state_eef:
      provider: urdf_fk
      pose_frame: galaxea_recorded_shared
      position_unit: m
      recorded: true
      joint_bindings:
        left_arm_joint1:  {signal: arm_joints, dimension: "left_arm./hdas/feedback_arm_left.position[0]", unit: rad}
        left_arm_joint2:  {signal: arm_joints, dimension: "left_arm./hdas/feedback_arm_left.position[1]", unit: rad}
        left_arm_joint3:  {signal: arm_joints, dimension: "left_arm./hdas/feedback_arm_left.position[2]", unit: rad}
        left_arm_joint4:  {signal: arm_joints, dimension: "left_arm./hdas/feedback_arm_left.position[3]", unit: rad}
        left_arm_joint5:  {signal: arm_joints, dimension: "left_arm./hdas/feedback_arm_left.position[4]", unit: rad}
        left_arm_joint6:  {signal: arm_joints, dimension: "left_arm./hdas/feedback_arm_left.position[5]", unit: rad}
        right_arm_joint1: {signal: arm_joints, dimension: "right_arm./hdas/feedback_arm_right.position[0]", unit: rad}
        right_arm_joint2: {signal: arm_joints, dimension: "right_arm./hdas/feedback_arm_right.position[1]", unit: rad}
        right_arm_joint3: {signal: arm_joints, dimension: "right_arm./hdas/feedback_arm_right.position[2]", unit: rad}
        right_arm_joint4: {signal: arm_joints, dimension: "right_arm./hdas/feedback_arm_right.position[3]", unit: rad}
        right_arm_joint5: {signal: arm_joints, dimension: "right_arm./hdas/feedback_arm_right.position[4]", unit: rad}
        right_arm_joint6: {signal: arm_joints, dimension: "right_arm./hdas/feedback_arm_right.position[5]", unit: rad}
        # 躯干在 base_link→臂 的链上,必须绑定;URDF 只有 3 个 torso joint
        torso_joint1: {signal: torso, dimension: "torso./hdas/feedback_torso.position[0]", unit: rad}
        torso_joint2: {signal: torso, dimension: "torso./hdas/feedback_torso.position[1]", unit: rad}
        torso_joint3: {signal: torso, dimension: "torso./hdas/feedback_torso.position[2]", unit: rad}
      position_signal: ee_pose
      position_dimensions:
        - "left_ee_pose./motion_control/pose_ee_arm_left.pose.position.x"
        - "left_ee_pose./motion_control/pose_ee_arm_left.pose.position.y"
        - "left_ee_pose./motion_control/pose_ee_arm_left.pose.position.z"
        - "right_ee_pose./motion_control/pose_ee_arm_right.pose.position.x"
        - "right_ee_pose./motion_control/pose_ee_arm_right.pose.position.y"
        - "right_ee_pose./motion_control/pose_ee_arm_right.pose.position.z"
      quaternion_signal: ee_pose
      quaternion_order: xyzw            # names 明确是 orientation.x/y/z/w
      quaternion_dimensions:
        - "left_ee_pose./motion_control/pose_ee_arm_left.pose.orientation.x"
        - "left_ee_pose./motion_control/pose_ee_arm_left.pose.orientation.y"
        - "left_ee_pose./motion_control/pose_ee_arm_left.pose.orientation.z"
        - "left_ee_pose./motion_control/pose_ee_arm_left.pose.orientation.w"
        - "right_ee_pose./motion_control/pose_ee_arm_right.pose.orientation.x"
        - "right_ee_pose./motion_control/pose_ee_arm_right.pose.orientation.y"
        - "right_ee_pose./motion_control/pose_ee_arm_right.pose.orientation.z"
        - "right_ee_pose./motion_control/pose_ee_arm_right.pose.orientation.w"
      effectors:
        - {name: left,  base_link: base_link, target_link: left_arm_link6}
        - {name: right, base_link: base_link, target_link: right_arm_link6}
    # action 没有记录 EEF,也没有躯干列;躯干仍用 state torso。
    action_eef:
      provider: urdf_fk
      pose_frame: galaxea_recorded_shared
      position_unit: m
      recorded: false
      joint_bindings:
        left_arm_joint1:  {signal: action_arms, dimension: "left_arm./motion_target/target_joint_state_arm_left.position[0]", unit: rad}
        left_arm_joint2:  {signal: action_arms, dimension: "left_arm./motion_target/target_joint_state_arm_left.position[1]", unit: rad}
        left_arm_joint3:  {signal: action_arms, dimension: "left_arm./motion_target/target_joint_state_arm_left.position[2]", unit: rad}
        left_arm_joint4:  {signal: action_arms, dimension: "left_arm./motion_target/target_joint_state_arm_left.position[3]", unit: rad}
        left_arm_joint5:  {signal: action_arms, dimension: "left_arm./motion_target/target_joint_state_arm_left.position[4]", unit: rad}
        left_arm_joint6:  {signal: action_arms, dimension: "left_arm./motion_target/target_joint_state_arm_left.position[5]", unit: rad}
        right_arm_joint1: {signal: action_arms, dimension: "right_arm./motion_target/target_joint_state_arm_right.position[0]", unit: rad}
        right_arm_joint2: {signal: action_arms, dimension: "right_arm./motion_target/target_joint_state_arm_right.position[1]", unit: rad}
        right_arm_joint3: {signal: action_arms, dimension: "right_arm./motion_target/target_joint_state_arm_right.position[2]", unit: rad}
        right_arm_joint4: {signal: action_arms, dimension: "right_arm./motion_target/target_joint_state_arm_right.position[3]", unit: rad}
        right_arm_joint5: {signal: action_arms, dimension: "right_arm./motion_target/target_joint_state_arm_right.position[4]", unit: rad}
        right_arm_joint6: {signal: action_arms, dimension: "right_arm./motion_target/target_joint_state_arm_right.position[5]", unit: rad}
        torso_joint1: {signal: torso, dimension: "torso./hdas/feedback_torso.position[0]", unit: rad}
        torso_joint2: {signal: torso, dimension: "torso./hdas/feedback_torso.position[1]", unit: rad}
        torso_joint3: {signal: torso, dimension: "torso./hdas/feedback_torso.position[2]", unit: rad}
      effectors:
        - {name: left, base_link: base_link, target_link: left_arm_link6}
        - {name: right, base_link: base_link, target_link: right_arm_link6}# ---- Per-check thresholds (tuned for 14-dim bimanual, float64) ----
advanced:
  pose_correction:
    enabled: true
  fk_consistency:
    enabled: true
    effect: annotate_only
    profile:
      schema_version: robot_curation.fk_consistency_profile.v2
      profile_id: galaxea.r1lite.s4.engineering.v1
      profile_version: "2"
      pose_key: state_eef
      pose_frame: galaxea_recorded_shared
      effector_names: [left, right]
      # 工程验证阈值;未经数据集级标定,不改变发布决策。
      translation_error_m_max: [0.03, 0.03]
      rotation_error_deg_max: [5.0, 5.0]

开启后跑完任务,输出的report目录下的summary.json中将写入transform_matrices数组信息。假如是双臂机器人,内容将包括T_base_to_data_left,T_base_to_data_right,T_eef_offset_left,T_eef_offset_right

基坐标系对齐

配置支持运行基坐标系对齐。

约束条件:

  • 环境变量mode=filter。
  • 依赖fk_consistency校验环节生成的产物(report/summary.json)中的transform_matrices数组信息。
  • 描述文件manifest中fk_consistency.enabled=true。
  • 描述文件manifest中轴标定信息forward_axis_in_base和up_axis_in_base为必填参数。

以下为agibot_a2d支持pose_correction的描述文件manifest示例:

# Galaxea (R1-Lite, bimanual) curation config (merged on top of global.yaml).
# Sampling: Galaxea frames are large (720x1280), so decode fewer per episode
# than the global default of 300 to keep video checks affordable.
paths:
  # Local dataset root — the directory IS the LeRobot v3 dataset root
  # (contains data/, meta/, videos/ directly).
  source_root: "D:/downloads/9e14315e5b23431e86821fb0512e2052"

sampling:
  video_max_frames: 200


model:
  path: service/app/configs/robot_urdf/agibot-G1_120s.urdf
  id: G1_120s.urdf
  sha256: fdab90abab488633ebdc3a474d32c431c6c70a8504501f62dec32845173e6c91
  # 轴标定,必填
  forward_axis_in_base: [1, 0, 0]
  up_axis_in_base: [0, 0, 1]


dataset:
  name: guodi_a2d
  repo_id: ""
  subdir: ""
  robot_type: a2d

  # ---- Field mapping ----
  # Galaxea has NO unified action/state columns and NO language_instruction.
  # We concatenate component fields into EpisodeView state/action arrays.
  state_fields: [observation.states.joint.position, observation.states.waist.position]
  action_fields: [actions.joint.position, actions.waist.position]
  language_fields: []                   # no language columns in parquet
  # Task text comes from the episode 'tasks' list / task_index.
  task_from_task_index: true

  signal_fields:
    arm_joints: [observation.states.joint.position]
    waist: [observation.states.waist.position]
    recorded_flange_position: [observation.source_eef.position]
    recorded_flange_orientation: [observation.source_eef.orientation]
    action_eef_pos: [ actions.end.position ]
    action_eef_quat: [ actions.end.orientation ]
    action_joints: [ actions.joint.position ]
    action_waist: [ actions.waist.position ]
  kinematic_poses:
    state_eef:
      provider: urdf_fk
      pose_frame: agibot_recorded_shared
      position_unit: m
      recorded: true
      joint_bindings:
        idx21_arm_l_joint1: {signal: arm_joints, dimension: position.left_arm_0, unit: rad}
        idx22_arm_l_joint2: {signal: arm_joints, dimension: position.left_arm_1, unit: rad}
        idx23_arm_l_joint3: {signal: arm_joints, dimension: position.left_arm_2, unit: rad}
        idx24_arm_l_joint4: {signal: arm_joints, dimension: position.left_arm_3, unit: rad}
        idx25_arm_l_joint5: {signal: arm_joints, dimension: position.left_arm_4, unit: rad}
        idx26_arm_l_joint6: {signal: arm_joints, dimension: position.left_arm_5, unit: rad}
        idx27_arm_l_joint7: {signal: arm_joints, dimension: position.left_arm_6, unit: rad}
        idx61_arm_r_joint1: {signal: arm_joints, dimension: position.right_arm_0, unit: rad}
        idx62_arm_r_joint2: {signal: arm_joints, dimension: position.right_arm_1, unit: rad}
        idx63_arm_r_joint3: {signal: arm_joints, dimension: position.right_arm_2, unit: rad}
        idx64_arm_r_joint4: {signal: arm_joints, dimension: position.right_arm_3, unit: rad}
        idx65_arm_r_joint5: {signal: arm_joints, dimension: position.right_arm_4, unit: rad}
        idx66_arm_r_joint6: {signal: arm_joints, dimension: position.right_arm_5, unit: rad}
        idx67_arm_r_joint7: {signal: arm_joints, dimension: position.right_arm_6, unit: rad}
        # 全量字段顺序为 pitch/lift;URDF joint1 是升降,joint2 是俯仰。
        idx01_body_joint1: {signal: waist, dimension: position.lift, unit: m}
        idx02_body_joint2: {signal: waist, dimension: position.pitch, unit: rad}
      position_signal: recorded_flange_position
      # names=null 的二维数组按行展开,d0..2 为左臂 xyz,d3..5 为右臂 xyz。
      position_dimensions:
        [observation.source_eef.position.d0, observation.source_eef.position.d1,
         observation.source_eef.position.d2, observation.source_eef.position.d3,
         observation.source_eef.position.d4, observation.source_eef.position.d5]
      quaternion_signal: recorded_flange_orientation
      quaternion_order: xyzw
      quaternion_dimensions:
        [observation.source_eef.orientation.d0, observation.source_eef.orientation.d1,
         observation.source_eef.orientation.d2, observation.source_eef.orientation.d3,
         observation.source_eef.orientation.d4, observation.source_eef.orientation.d5,
         observation.source_eef.orientation.d6, observation.source_eef.orientation.d7]
      effectors:
        - {name: left, base_link: base_link, target_link: arm_l_end_link}
        - {name: right, base_link: base_link, target_link: arm_r_end_link}
         T_pose_from_model_base:
        [1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, -0.3, 0, 0, 0, 1]
    action_eef:
      provider: urdf_fk
      pose_frame: agibot_recorded_shared
      position_unit: m
      recorded: true
      joint_bindings:
        idx21_arm_l_joint1: { signal: action_joints, dimension: position.left_arm_0, unit: rad }
        idx22_arm_l_joint2: { signal: action_joints, dimension: position.left_arm_1, unit: rad }
        idx23_arm_l_joint3: { signal: action_joints, dimension: position.left_arm_2, unit: rad }
        idx24_arm_l_joint4: { signal: action_joints, dimension: position.left_arm_3, unit: rad }
        idx25_arm_l_joint5: { signal: action_joints, dimension: position.left_arm_4, unit: rad }
        idx26_arm_l_joint6: { signal: action_joints, dimension: position.left_arm_5, unit: rad }
        idx27_arm_l_joint7: { signal: action_joints, dimension: position.left_arm_6, unit: rad }
        idx61_arm_r_joint1: { signal: action_joints, dimension: position.right_arm_0, unit: rad }
        idx62_arm_r_joint2: { signal: action_joints, dimension: position.right_arm_1, unit: rad }
        idx63_arm_r_joint3: { signal: action_joints, dimension: position.right_arm_2, unit: rad }
        idx64_arm_r_joint4: { signal: action_joints, dimension: position.right_arm_3, unit: rad }
        idx65_arm_r_joint5: { signal: action_joints, dimension: position.right_arm_4, unit: rad }
        idx66_arm_r_joint6: { signal: action_joints, dimension: position.right_arm_5, unit: rad }
        idx67_arm_r_joint7: { signal: action_joints, dimension: position.right_arm_6, unit: rad }
        idx01_body_joint1: { signal: action_waist, dimension: position.lift, unit: m }
        idx02_body_joint2: { signal: action_waist, dimension: position.pitch, unit: rad }
      position_signal: action_eef_pos
      position_dimensions:
        [ position.left_xyz.d0, position.left_xyz.d1, position.left_xyz.d2,
          position.right_xyz.d0, position.right_xyz.d1, position.right_xyz.d2 ]
      quaternion_signal: action_eef_quat
      quaternion_order: xyzw
      quaternion_dimensions:
        [ orientation.left_xyzw.d0, orientation.left_xyzw.d1, orientation.left_xyzw.d2, orientation.left_xyzw.d3,
          orientation.right_xyzw.d0, orientation.right_xyzw.d1, orientation.right_xyzw.d2, orientation.right_xyzw.d3 ]
      effectors:
        - { name: left, base_link: base_link, target_link: arm_l_end_link }
        - { name: right, base_link: base_link, target_link: arm_r_end_link }
      T_pose_from_model_base:
        [ 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, -0.3, 0, 0, 0, 1 ]

# ---- Per-check thresholds (tuned for 12+2 dim bimanual, float64) ----
checks:
  nan_inf:
    enabled: true
    arrays: [state, action]

  language:
    enabled: false
    # Since Galaxea has no language columns, the check runs against the
    # resolved task text (EpisodeView.task) when task_from_task_index=true.
    min_chars: 1
    max_chars: 1000
    visible_ratio_min: 0.8
    repeated_char_run_max: 40

  episode_length:
    enabled: true
    abs_min: 5
    abs_max: 5000
    low_percentile: 1
    high_percentile: 99

  video_quality:
    enabled: true
    black_mean_max: 15.0
    black_lowvar_ratio: 0.95
    # NOTE: Laplacian variance scales with resolution. Galaxea frames are
    # 720x1280 and their blur scores are ~3-7x lower than DROID's 180x320
    # frames for similar sharpness, so use a lower threshold here.
    blur_laplacian_min: 30.0

    blur_metric: relative           # ← 加这行,默认 laplacian
    blur_relative_min: 1.25         # ← 比值下限
    blur_signal_min: 40.0

  sudden_change:
    enabled: true
    smoothing: savgol
    savgol_window: 7
    savgol_poly: 2
    moving_window: 5
    residual_threshold: 0.2
    accel_threshold: 1.0
    jerk_threshold: 1.5
    require_accel_or_jerk: true
    # exclude_gripper (optional, default: false):
    #   When set to true, gripper action dimensions are skipped during this
    #   check. Enabling it requires `gripper_dims` to be configured; if
    #   `gripper_dims` is empty or all listed dims are out of range, the
    #   exclusion has no effect.
    exclude_gripper: true
    # gripper dims are indices 0 and 1 in the concatenated action
    gripper_dims: [14, 15]

  extreme_value:
    enabled: true
    lower_q: 0.01
    upper_q: 0.99
    alpha: 3.0
    # gripper dims are indices 12 and 13 in the concatenated action (6+6+1+1)
    gripper_dims: [14, 15]
    gripper_unique_min: 2

  state_action_alignment:
    enabled: true
    smoothing: savgol
    savgol_window: 7
    savgol_poly: 2
    max_lag: 5
    direction_agree_min: 0.6


advanced:
  pose_correction:
    enabled: true
  fk_consistency:
    enabled: true
    effect: annotate_only
    profile:
      schema_version: robot_curation.fk_consistency_profile.v2
      profile_id: agibot.extended-v1.g1.s4.engineering
      profile_version: "2"
      pose_key: state_eef
      pose_frame: agibot_recorded_shared
      effector_names: [left, right]
      # 工程验证阈值;未经数据集级标定,不改变发布决策。
      translation_error_m_max: [0.03, 0.03]
      rotation_error_deg_max: [5.0, 5.0]

输出数据集的data/parquet中将写入校准信息,meta/info.json中将写入对应的元信息。

    "observation.states.left_eef_calibration": {
      "dtype": "float32",
      "shape": [
        7
      ],
      "names": [
        "x",
        "y",
        "z",
        "qx",
        "qy",
        "qz",
        "qw"
      ],
      "fps": 30.0
    },
    "observation.states.right_eef_calibration": {
      "dtype": "float32",
      "shape": [
        7
      ],
      "names": [
        "x",
        "y",
        "z",
        "qx",
        "qy",
        "qz",
        "qw"
      ],
      "fps": 30.0
    },
    "actions.left_eef_calibration": {
      "dtype": "float32",
      "shape": [
        7
      ],
      "names": [
        "x",
        "y",
        "z",
        "qx",
        "qy",
        "qz",
        "qw"
      ],
      "fps": 30.0
    },
    "actions.right_eef_calibration": {
      "dtype": "float32",
      "shape": [
        7
      ],
      "names": [
        "x",
        "y",
        "z",
        "qx",
        "qy",
        "qz",
        "qw"
      ],

相关文档