README.md

metadata

license: cc-by-nc-sa-4.0

M7 pick place example

Dataset Description

This dataset contains multimodal robotic manipulation trajectories of an M7 Robot equipped with the xhand_norm dexterous hand. The recorded task is a fundamental tabletop Pick-and-Place sequence involving two distinct objects.

The dataset is formatted for seamless integration with the LeRobot library, making it ideal for training Vision-Language-Action (VLA) models.

---

Dataset Structure

Directory Layout

Pick_Apple_Mango/
├── 2031605/              # Split 1 (491 episodes, 118,555 frames)
│   ├── data/
│   │   └── chunk-000/
│   │       ├── episode_000000.parquet
│   │       ├── episode_000001.parquet
│   │       └── ...
│   ├── meta/
│   │   ├── episodes.jsonl      # episode length & task descriptions
│   │   ├── info.json           # field definitions, robot type, etc.
│   │   ├── joint_indexes.json  # semantic names for action/state dimensions
│   │   └── tasks.jsonl         # 16 natural-language task instructions
│   └── videos/
│       └── ...
├── 2031607/              # Split 2 (255 episodes, 46,742 frames)
│   └── ...
├── 2031704/             # Split 3 (720 episodes, 121,158 frames)
│   └── ...
└── .cache/              # LeRobot internal cache (safe to ignore)

Data Format

Each record (one row in a parquet file) follows the LeRobot format:

Field	Type	Shape	Description
observation.images.cam_high	video (h264)	[480, 848, 3]	Overhead camera, 848×480, 20fps
observation.images.cam_left	video (h264)	[480, 640, 3]	Left shoulder camera, 640×480, 20fps
observation.images.cam_right	video (h264)	[480, 640, 3]	Right shoulder camera, 640×480, 20fps
observation.state	float32	[57]	Full robot state (arms + end-effectors + hands)
action	float32	[38]	Dual-arm end-effector + dexterous hand actions
episode_index	int64	[1]	Episode index
frame_index	int64	[1]	Frame index within the episode
task_index	int64	[1]	Task index (0–15)
timestamp	float32	[1]	Timestamp (seconds)

Note: Actions are expressed as absolute end-effector pose (xyz + quaternion) combined with dexterous hand joint angles. See meta/joint_indexes.json for full semantic labels.

---

Dataset Statistics

⚠️ All three splits combined

Statistic	Value
Total Episodes	1,466
Total Frames	286,455
Total Videos	4,398
Robot Type	M7_XHAND1 (dual-arm + bi-manual hands)
Average Episode Length	~195 frames
FPS	20 Hz
Collection Environment	Indoor tabletop (real-world)
Object Categories	Apple, Mango

Split Breakdown

Split	Episodes	Frames	Videos
2031605	491	118,555	1,473
2031607	255	46,742	765
2031704	720	121,158	2,160

Action / State Dimensions

Field	Dimensions	Composition
action	38	Right end-effector (7: xyz + quat(xyzw)) + Right hand (12) + Left end-effector (7) + Left hand (12)
observation.state	57	Right arm joints (7) + Left arm joints (7) + Dual end-effector poses (14) + Dual hands (24) + Waist/Neck (5)

---

Camera Configuration

Camera	Resolution	FPS	Codec	View
cam_high	848×480	20	h264	Overhead — full tabletop view
cam_left	640×480	20	h264	Left hand perspective
cam_right	640×480	20	h264	Right hand perspective

---

Robot System

• Model: M7_XHAND1 (dual-arm collaborative robot)
• Arms: 7 DoF each (shoulder_pitch, shoulder_roll, arm_yaw, elbow_pitch, elbow_yaw, wrist_pitch, wrist_roll)
• Dexterous Hands: 12 DoF per hand (thumb_bend, thumb_rota1, thumb_rota2, index_bend, index_joint1, index_joint2, mid_joint1, mid_joint2, ring_joint1, ring_joint2, pinky_joint1, pinky_joint2; left hand has an additional ring_joint2)
• Additional Joints: Waist 3 DoF + Neck 2 DoF
• End-Effectors: 7 DoF (position xyz + quaternion orientation (xyzw))
• Dataset Codebase Version: LeRobot v2.0

---

Dataset Card Authors