遇见数据集

robomimic-can-paired-lerobot

收藏
Hugging Face2026-08-28 更新2026-08-28 收录
官方服务:

资源简介:

robomimic can-paired数据集,转换为LeRobot v3.0格式,并带有成功/失败标签。原始robomimic can-paired数据集包含200个遥操作演示,执行robosuite的PickPlaceCan任务(Panda机器人):100个配对任务初始化,每个配对包含一个成功演示(罐子被拿起并放入正确箱子)和一个失败演示(罐子被拿起并扔出机器人工作空间)。摄像头观测(agentview、robot0_eye_in_hand腕部摄像头)使用MuJoCo/robosuite v1.5.1在256x256分辨率下从原始demo_v15.hdf5状态重新渲染。成功/失败标签:每帧有布尔特征next.success(在episode内恒定);每个episode的任务字符串以(success)或(failure)结尾,可通过任务元数据过滤,恰好100个成功/100个失败episode。标签衍生自稀疏任务奖励(success = max reward > 0)并与官方robomimic low_dim_v15.hdf5奖励交叉验证。特征包括:observation.images.agentview和wrist摄像头(256x256x3视频,20 fps)、observation.state(9维:末端执行器位置3,四元数4,夹爪位置2)、observation.environment_state(14维:物体罐子的位姿特征)、action(7维:OSC增量位姿6 + 夹爪1)、next.reward、next.done、next.success。来源:原始数据来自amandlek/robomimic数据集v1.5/can/paired;robomimic论文:Mandlekar等人,CoRL 2021。主页:https://robomimic.github.io/,论文:https://arxiv.org/abs/2108.03298,许可证:MIT。

The robomimic can-paired dataset, converted to LeRobot v3.0 format with success/failure labels. The original robomimic can-paired dataset contains 200 teleoperated demonstrations performing the robosuite PickPlaceCan task (Panda robot): 100 paired task initializations, each pair contains a successful demonstration (can picked up and placed into correct bin) and a failed demonstration (can picked up and thrown out of robot workspace). Camera observations (agentview, robot0_eye_in_hand wrist camera) are re-rendered at 256x256 resolution from the original demo_v15.hdf5 states using MuJoCo/robosuite v1.5.1. Success/failure labels: boolean feature next.success per frame (constant within episode); each episodes task string ends with (success) or (failure), can be filtered by task metadata, exactly 100 success/100 failure episodes. Labels derived from sparse task rewards (success = max reward > 0) and cross-validated with official robomimic low_dim_v15.hdf5 rewards. Features include: observation.images.agentview and wrist cameras (256x256x3 video, 20 fps), observation.state (9-dim: end-effector position 3, quaternion 4, gripper position 2), observation.environment_state (14-dim: object can pose features), action (7-dim: OSC delta pose 6 + gripper 1), next.reward, next.done, next.success. Source: original data from amandlek/robomimic dataset v1.5/can/paired; robomimic paper: Mandlekar et al., CoRL 2021. Homepage: https://robomimic.github.io/, Paper: https://arxiv.org/abs/2108.03298, License: MIT.

提供机构:
geonmin-kim
创建时间:
2026-08-28
原始信息汇总

数据集概述:robomimic-can-paired-lerobot

基本信息

数据集内容

该数据集源自robomimic的can-paired数据集,包含200个遥操作演示,对应robosuite中的PickPlaceCan任务(Panda机器人),其中:

  • 100个成功演示:罐子被拾取并放入正确箱子
  • 100个失败演示:罐子被拾取但被抛出机器人工作空间

每个任务初始化包含一对成功/失败演示。摄像机观察(agentviewrobot0_eye_in_hand腕部相机)使用MuJoCo/robosuite v1.5.1在离屏条件下以256x256分辨率重新渲染,数据来自原始的demo_v15.hdf5状态。

成功/失败标签

  • 逐帧标签:布尔特征next.success(在单个episode内保持恒定)
  • 逐episode标签:任务字符串以(success)(failure)结尾,可通过任务元数据过滤episode
  • 标签来源:从稀疏任务奖励(success = max reward > 0)推导,并与官方robomimic low_dim_v15.hdf5奖励交叉验证

特征说明

特征 维度 描述
observation.images.agentview 256x256x3 顶部视角视频 @ 20 fps
observation.images.wrist 256x256x3 腕部相机视频 @ 20 fps
observation.state 9 末端执行器位置(3)、四元数(4)、夹爪位置(2)
observation.environment_state 14 物体(罐子)位姿特征
action 7 OSC增量位姿(6) + 夹爪控制(1)
next.reward 1 奖励值
next.done 1 完成标志
next.success 1 成功标志

数据规模

  • 总episode数: 200
  • 总帧数: 19,795
  • 总任务数: 2
  • 数据文件大小: 约100 MB(Parquet格式)
  • 视频文件大小: 约200 MB(AV1编码,MP4格式)
  • 帧率: 20 fps
  • 数据划分: 仅训练集(0:200)

零样本检查点

该数据集为偏好学习任务指定了零样本策略:lerobot/smolvla_base(约450M参数,VLM骨干+流匹配动作专家)。使用时需注意:

  • 需将本数据集的observation.state(9维)和action(7维)映射到SmolVLA的填充状态/动作约定
  • 该模型主要在SO-10x真实机器人社区数据上预训练,预期在robosuite PickPlaceCan任务上的零样本闭环成功率较低
  • 官方实测零样本闭环成功率待定(TBD)

数据来源

  • 原始数据来自amandlek/robomimicv1.5/can/paired子集
  • 引用参考:Mandlekar et al., "What Matters in Learning from Offline Human Demonstrations for Robot Manipulation", CoRL 2021
二维码
社区交流群
二维码
科研交流群
商业服务