xarm-hang-multimug-v0

A multi-task teleoperation dataset for hanging mugs of different colors on a mug tree, collected on an xArm7 robot arm via phone teleoperation.

Overview

Field	Value
Robot	xArm7
Tasks	3
Episodes	500
Total frames	305,540
Total duration	169.7 min (2 h 50 min)
FPS	30
Format	LeRobot v3.0

Tasks

task_index	Description	Episodes	Duration
0	Hang the red mug on the mug tree	165	~58.0 min
1	Hang the white mug on the mug tree	165	~54.3 min
2	Hang the blue mug on the mug tree	170	~57.5 min

Episode Length

Stat	Frames	Seconds
Min	346	11.5
Median	608	20.3
Mean	611	20.4
Max	1002	33.4

Observation Space

State (observation.state) — float32[15]

Index	Name	Description
0–2	tcp.x/y/z	TCP position (mm)
3–6	tcp.qw/qx/qy/qz	TCP orientation as quaternion (scalar-first, canonical w ≥ 0)
7–13	joint.0–joint.6	Joint angles (degrees)
14	gripper	Gripper state

Cameras

Key	Resolution	FPS	Codec	Description
observation.images.ego	320×240	30	h264	Third-person egocentric view (downsampled 2× from 640×480)
observation.images.wrist	320×240	30	h264	Wrist-mounted camera (downsampled 2× from 640×480)

Action Space

Action (action) — float32[8]

Index	Name	Description
0–2	tcp.x/y/z	Target TCP position (mm)
3–6	tcp.qw/qx/qy/qz	Target TCP orientation as quaternion (scalar-first)
7	gripper	Target gripper command

Dataset Structure

xarm-hang-multimug-v0/
├── data/
│   └── chunk-000/
│       └── file-000.parquet          # All frames (305,540 rows)
├── videos/
│   └── chunk-000/
│       ├── observation.images.ego/   # 2 × file-NNN.mp4 (packed, ≤200 MB each)
│       └── observation.images.wrist/ # 2 × file-NNN.mp4 (packed, ≤200 MB each)
└── meta/
    ├── info.json
    ├── stats.json
    ├── tasks.parquet
    └── episodes/
        └── chunk-000/
            └── file-000.parquet      # Per-episode metadata and stats

Videos were recorded one MP4 per episode at 640×480, then packed into multi-episode files and downsampled 2× via teleoperation/combined_videos.py. Per-episode timestamp ranges are recorded in meta/episodes/.../file-000.parquet under videos/<key>/from_timestamp and to_timestamp.