mimicgen_hammer_cleanup_d1_224x224_mtdit_flow_35k_failure
收藏资源简介:
该数据集是使用LeRobot创建的机器人模仿学习数据集,包含MimicGen模拟环境中锤子清理任务的失败轨迹。数据集由100个episodes组成,共50000帧,以20fps采样,包含两个摄像头视角的图像观测(agentview和eye_in_hand,分辨率224x224)以及丰富的状态信息,如机器人关节位置、速度、末端执行器位置、姿态、力/力矩等。该数据集专为研究机器人失败场景下的模仿学习而设计,可用于分析失败模式、提高策略鲁棒性等任务。
This dataset is a robot imitation learning dataset created using LeRobot, containing failure trajectories of the hammer cleaning task in the MimicGen simulation environment. It consists of 100 episodes totaling 50,000 frames sampled at 20fps, with image observations from two camera views (agentview and eye_in_hand, resolution 224x224) and rich state information such as robot joint positions, velocities, end-effector positions, orientations, forces/torques, etc. The dataset is specifically designed for studying imitation learning under robot failure scenarios, and can be used for analyzing failure modes, improving policy robustness, etc.
数据集概述:mimicgen_hammer_cleanup_d1_224x224_mtdit_flow_35k_failure
基本信息
- 许可证: cc-by-4.0
- 任务类别: 机器人技术(robotics)
- 标签: LeRobot、mimicgen、robosuite、v3.0、rollout、failure、dice-rl
- 机器人类型: Panda
数据集内容
该数据集是一个基于LeRobot框架创建的机器人操作数据集,专门用于锤子清理(hammer cleanup)任务。数据集包含了100个回合(episodes),总帧数为50000帧,采样频率为20 FPS,仅包含1个任务。
数据结构
特征信息
| 特征 | 数据类型 | 形状 |
|---|---|---|
| action | float32 | 7 |
| observation.state | float32 | 37 |
| observation.sim_state | float32 | 34 |
| reward | float32 | 1 |
| done | bool | 1 |
| success | bool | 1 |
| is_first / is_last / is_terminal | bool | 1 |
| source_demo_index | int64 | 1 |
| observation.images.agentview | video | 224×224×3 |
| observation.images.eye_in_hand | video | 224×224×3 |
| observation.object | float32 | 28 |
| observation.robot0_contact | float32 | 1 |
| observation.robot0_eef_force_norm | float32 | 1 |
| observation.robot0_eef_pos | float32 | 3 |
| observation.robot0_eef_quat | float32 | 4 |
| observation.robot0_eef_vel_ang | float32 | 3 |
| observation.robot0_eef_vel_lin | float32 | 3 |
| observation.robot0_gripper_qpos | float32 | 2 |
| observation.robot0_gripper_qvel | float32 | 2 |
| observation.robot0_joint_pos / cos / sin / vel | float32 | 7 |
| timestamp | float32 | 1 |
| frame_index / episode_index / index / task_index | int64 | 1 |
图像信息
- 两个摄像头视角:agentview和eye_in_hand
- 分辨率:224×224,3通道
- 视频编码:AV1,20 FPS,无音频
数据划分与存储
- 训练集划分: train 0:100
- 数据文件路径: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
- 视频文件路径: videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
- 数据大小: 数据文件约100 MB,视频文件约200 MB
数据集用途
该数据集标记为“failure”,表明其记录了机器人执行锤子清理任务时的失败轨迹数据,可能用于研究机器人操作中的故障模式分析和鲁棒性训练,与dice-rl技术相关。



