MetaFine_CoRL26
收藏资源简介:
该数据集是 CoRL demos 数据集,用于 MetaFine π0 基线模型的训练与评估,涵盖五个机器人操作任务(T1–T5)。每个任务包含多种变体,并具有统一的观测与控制空间。数据集包含 RGB 图像(来自 base 和 hand 摄像头,分辨率 512×512,视野约 70° 或 57°)以及 9 自由度关节位置增量控制信号(pd_joint_delta_pos),以 30 FPS 采集。具体任务如下:T1 grasp_part(抓取部件,200 个混合片段,30,670 帧,变体 cap/body/mixed);T2 grasp_move_mug(抓取并移动杯子,300 个混合片段,39,395 帧,变体 left/right/forward/mixed);T3 toggle_switch_table(切换开关,200 个混合片段,20,567 帧,变体 red/blue/mixed);T4 put_blocks_into_boxes(将方块放入盒子,300 个混合片段,156,151 帧,变体 red/blue/green/mixed);T5 insert_letter(插入字母,400 个混合片段,115,894 帧,变体 C/o/R/L/mixed)。每个变体的片段数均为 100(混合任务合并了所有变体)。数据以 LeRobot 格式组织,包含运动规划源轨迹(.h5/.json)、RGB 回放文件(.rgb.pd_joint_delta_pos.physx_cpu.h5/.json)以及 HuggingFace LeRobot 数据集目录(meta/、data/、videos/)。部分任务(T4、T5)还包含可选的 ManiSkill 原始片段转储。该数据集适用于机器人操作领域的模仿学习、行为克隆等任务。
This dataset is the CoRL demos dataset for training and evaluating the MetaFine π0 baseline model, covering five robot manipulation tasks (T1–T5). Each task includes multiple variants and has a unified observation and control space. The dataset contains RGB images (from base and hand cameras, resolution 512×512, field of view approximately 70° or 57°) and 9-DoF joint position delta control signals (pd_joint_delta_pos), collected at 30 FPS. The specific tasks are: T1 grasp_part (200 mixed episodes, 30,670 frames, variants cap/body/mixed); T2 grasp_move_mug (300 mixed episodes, 39,395 frames, variants left/right/forward/mixed); T3 toggle_switch_table (200 mixed episodes, 20,567 frames, variants red/blue/mixed); T4 put_blocks_into_boxes (300 mixed episodes, 156,151 frames, variants red/blue/green/mixed); T5 insert_letter (400 mixed episodes, 115,894 frames, variants C/o/R/L/mixed). Each variant has 100 episodes (mixed tasks combine all variants). Data is organized in LeRobot format, including motion planning source trajectories (.h5/.json), RGB replay files (.rgb.pd_joint_delta_pos.physx_cpu.h5/.json), and HuggingFace LeRobot dataset directories (meta/, data/, videos/). Some tasks (T4, T5) also include optional ManiSkill raw episode dumps. This dataset is suitable for imitation learning, behavior cloning, and other tasks in robot manipulation.
数据集概述:MetaFine_CoRL26
该数据集是 MetaFine π0 基线模型(任务 T1–T5)的训练数据,用于机器人操作学习。
任务与数据规模
| 任务 | 变体 | 混合片段数 | 混合帧数 | 相机 | 状态维度 | 帧率 | 控制/观测 |
|---|---|---|---|---|---|---|---|
T1 grasp_part |
cap / body / mixed | 200 | 30,670 | base + hand, 512×512, FOV ≈ 70° | 9-DoF | 30 | pd_joint_delta_pos / rgb |
T2 grasp_move_mug |
left / right / forward / mixed | 300 | 39,395 | base + hand, 512×512, FOV ≈ 70° | 9-DoF | 30 | pd_joint_delta_pos / rgb |
T3 toggle_switch_table |
red / blue / mixed | 200 | 20,567 | base + hand, 512×512, FOV ≈ 70° | 9-DoF | 30 | pd_joint_delta_pos / rgb |
T4 put_blocks_into_boxes |
red / blue / green / mixed | 300 | 156,151 | base + hand, 512×512, FOV ≈ 57° | 9-DoF | 30 | pd_joint_delta_pos / rgb |
T5 insert_letter |
C / o / R / L / mixed | 400 | 115,894 | base + hand, 512×512, FOV ≈ 70° | 9-DoF | 30 | pd_joint_delta_pos / rgb |
每个变体的片段数均为 100(T5 的每个字母变体亦为 100),混合变体为各变体的合并集。
对应训练检查点
| LeRobot 数据集 | 训练检查点 |
|---|---|
grasp_part/mixed/lerobot |
outputs/pi0_grasp_mixed_f70/checkpoints/030000/pretrained_model |
grasp_move_mug/mixed/lerobot |
outputs/pi0_grasp_move_mug_mixed/checkpoints/030000/pretrained_model |
toggle_switch_table/mixed/lerobot |
outputs/pi0_toggle_mixed/checkpoints/030000/pretrained_model |
put_blocks_into_boxes/mixed/lerobot |
outputs/pi0_put_blocks_mixed/checkpoints/030000/pretrained_model |
insert_letter/mixed/lerobot |
outputs/pi0_insert_letter_mixed/checkpoints/030000/pretrained_model |
目录结构
demos/CoRL/ grasp_part/ # 见 grasp_part/README.md grasp_move_mug/ # 见 grasp_move_mug/README.md (T2) toggle_switch_table/ # 见 toggle_switch_table/README.md (T3) put_blocks_into_boxes/ # 见 put_blocks_into_boxes/README.md (T4) insert_letter/ # 见 insert_letter/README.md (T5) lerobot/ # 旧符号链接 grasp_part_cap_n100_f70 → ../grasp_part/cap/lerobot grasp_part_body_n100_f70 → ../grasp_part/body/lerobot grasp_part_mixed_n200_f70 → ../grasp_part/mixed/lerobot
每个变体文件夹通常包含:
*.h5/*.json— 运动规划源轨迹*.rgb.pd_joint_delta_pos.physx_cpu.h5(+.json) — 用于 LeRobot 转换的 RGB 回放lerobot/— HuggingFace LeRobot 数据集(meta/、data/、videos/)raw/(T4 / T5)— 可选的 ManiSkill 原始片段导出
评估传感器配置与 RGB 回放元数据一致,见 eval/configs/*.yaml。




