MM-30

A real-world mobile manipulation dataset, released in a LeRobot v3.0 file layout.

Episode snapshots

Put fruit into a bowl	Hang cups on a rack	Sort blocks into cups
Stack bowls	Screw a cap onto a test tube	Open a microwave
Put fruit into a cabinet	Load a refrigerator	Fold a towel
Fold clothing	Open a bag and take out a toy	Sweep blocks into a dustpan

Overview

This dataset contains teleoperated demonstrations from a mobile dual-arm robot working in indoor tabletop, kitchen, office, and lab-style scenes. Every episode is a successful demonstration of one English instruction. People who appear in the recordings are company employees and consented to this public release.

The public release is the training set only. Five held-out demo directories from collection (demo1–demo5, 1,175 episodes) are not included. Two broken recordings of 5 and 15 frames were also dropped.

Episodes	2,443
Frames	3,383,139
Duration	31.33 hours
Frame rate	30 Hz
Tasks	68
Cameras	3 × RGB 640×480
Episode length	12.9 s – 138.6 s (median 43.1 s)
Language	English

中文简介

MM-30 是一套真实场景下的移动操作演示，按 LeRobot v3.0 的分片方式组织。共 2443 条成功演示、68 个英文任务、约 31.3 小时。画面中出现的人都是已授权的公司员工。采集时的 5 个 demo 目录（1,175 条）和 2 条残缺录制（5 帧、15 帧）没有放进这个公开发布版本。

Robot and sensors

• Mobile base. Command and measured velocity are base_twist = [vx, vy, wz] in m/s, m/s, rad/s.
• Vertical lift. lift_height is the column position in metres.
• Two 6-DoF arms. The arms are marked PIPER. Joints are radians. The canonical arm-joint slot is 7-wide; dimension 6 on each arm is padding and is always 0.
• Parallel gripper on each arm. Stored in [0, 1], where 0 is closed and 1 is open.
• Three RGB cameras, JPEG at capture and H.264 in this release, 640×480, 30 fps:
  • primary: head camera
  • wrist_left: left wrist
  • wrist_right: right wrist

Raw capture stored JPEG with red and blue swapped. The MP4 files in this release are color-corrected, so a normal player shows the original colors. Do not swap channels again.

What action means

The source logs have one proprioceptive stream and no separate command stream. In every parquet row, action is exactly equal to observation.state at that frame. Both are absolute quantities (joint position, end-effector pose, gripper opening, base velocity, lift height).

For delta or next-state training, compute the target from a future frame, for example state[t+k] - state[t] for joints and gripper, and a relative rigid transform for the end-effector. The absolute end-effector frame is the flange expressed at the arm base. It was not checked against a URDF, so prefer relative targets.

Directory layout

.
├── README.md
├── assets/
├── meta/
│   ├── info.json
│   ├── modality.json
│   ├── embodiment.json
│   ├── tasks.parquet
│   ├── task_instructions.json
│   └── episodes/chunk-000/file-000.parquet
├── data/chunk-{000,001,002}/file-{file_index:03d}.parquet
└── videos/{primary,wrist_left,wrist_right}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4

One parquet file and one MP4 per camera store one episode. Episodes are numbered episode_index = chunk_index * 1000 + file_index. Path templates in meta/info.json:

• data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
• videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4

codebase_version is v3.0. This is the sharded LeRobot v3 layout (chunk / file index, dataset_from_index, per-video timestamps). It is not the older v2.1 episode_000000.parquet naming used by some releases.

Parquet columns

Column	dtype	Shape	Meaning
observation.state	float32	(80,)	Absolute proprioception, canonical layout below
action	float32	(80,)	Same values as observation.state
timestamp	float32	scalar	Seconds from the start of the episode, frame_index / 30
frame_index	int64	scalar	0 … length−1 inside the episode
episode_index	int64	scalar	0 … 2442
index	int64	scalar	Global frame index, contiguous from 0
task_index	int64	scalar	Row in meta/tasks.parquet

Canonical 80-D layout

The same index map is used for observation.state and action. Dimensions not listed are reserved, stored as 0, and marked invalid in meta/embodiment.json.

Slice	Name	Unit	Notes
0:6	left_arm_joints	rad	6 joints. Index 6 is unused padding
7:10	left_eef_position	m	Flange xyz at the arm base
10:16	left_eef_rotation		Column-major rot6d, Zhou's first two columns
16:17	left_gripper	[0, 1]	0 closed, 1 open
29:35	right_arm_joints	rad	Index 35 is unused padding
36:39	right_eef_position	m
39:45	right_eef_rotation		Column-major rot6d
45:46	right_gripper	[0, 1]
58:61	base_twist	m/s, m/s, rad/s	[vx, vy, wz]
68:69	lift_height	m

End-effector rotation was converted from a unit quaternion in xyzw order (scipy / ROS). Gripper values were clipped to [0, 1]; source values were already inside that range.

Observed ranges on the released frames:

Signal	Min	Max
Left joints (rad)	−2.96	3.25
Left flange xyz (m)	x 0.02, y −0.50, z −0.22	x 0.61, y 0.39, z 0.47
Right flange xyz (m)	x −0.05, y −0.35, z −0.26	x 0.63, y 0.54, z 0.58
Gripper, both arms	0	1
Base vx, vy, wz	−0.62, −0.44, −0.79	0.62, 0.44, 0.79
Lift height (m)	0	0.16

Language

meta/tasks.parquet has one row per task: task_index and the primary instruction. That primary string is also stored on every episode in meta/episodes.

Collection wrote several paraphrases per task (usually 10, sometimes 6). Only the first is the official task_index string, so a loader that joins on task_index stays one-to-one. The full lists are in meta/task_instructions.json.

Tasks

Index	Episodes	Frames	Hours	Primary instruction
0	33	55,712	0.52	Screw the cap onto the test tube.
1	90	109,638	1.02	Hang the cups on the cup rack.
2	83	94,094	0.87	Gently push the file folders into the file organizer.
3	53	54,396	0.50	Remove the lid from the kettle.
4	52	55,514	0.51	Put the lid back on the kettle.
5	32	59,654	0.55	Pour the grains from the kettle into the bowl on the left.
6	43	59,514	0.55	Pour the grains from the kettle into the bowl on the right.
7	89	105,357	0.98	Open the microwave and remove the item inside.
8	91	108,998	1.01	Use the scoop to transfer some grains from the large bowl to the small bowl.
9	83	53,754	0.50	Set the tipped-over cup upright.
10	80	72,826	0.67	Move to the front of the table, then put the red block into the red cup and the yellow block into the yellow cup.
11	70	108,496	1.00	Put the umbrella from the table into the umbrella stand.
12	63	109,845	1.02	Put the fruit scattered on the table into the bowl.
13	88	109,876	1.02	Choose one umbrella at random from the umbrella stand and place it on the table.
14	47	83,475	0.77	Tighten the large bolt into the nut, then place the assembled pair on the table.
15	40	87,792	0.81	Unscrew the large bolt from the nut, then place both separated parts on the table.
16	26	55,100	0.51	Fold the towel on the table.
17	29	55,042	0.51	Unfold the towel on the table.
18	62	110,325	1.02	Open the bag and remove the toy from inside.
19	33	54,949	0.51	Fold the clothing on the table.
20	33	53,138	0.49	Unfold the clothing on the table.
21	90	108,534	1.00	Drive to the front of the table, pick up one pen with the right hand and the other with the left hand, then place both pens inside the pen holder.
22	66	111,643	1.03	Place the toy into the bag and zip the bag closed.
23	65	123,637	1.14	Find the table and sofa, then move the bag from the table to the sofa.
24	63	109,547	1.01	Insert the test tube from the table into the test tube rack.
25	84	111,234	1.03	Push the office chair underneath the desk.
26	61	109,752	1.02	Place the two water bottles from the table in front of their corresponding chairs.
27	56	110,408	1.02	Go to the table and stack the bowls.
28	80	114,283	1.06	Pick up the mouse from the table and place it on the mouse pad on the adjacent table.
29	47	111,840	1.04	Sweep the blocks into the dustpan, then pour them into the bowl on the adjacent table.
30	90	112,352	1.04	Unfold the napkin and cover the bread in the bread basket.
31	40	54,563	0.51	Go to the table and put the fruit from the tabletop into the cabinet.
32	10	11,606	0.11	Using one hand, lift the banana out of the basket on the current table and place it in the basket on the table behind you.
33	10	11,532	0.11	With one hand, pick up the bread slice from the basket and place it in the basket on the table behind you.
34	10	11,896	0.11	Pick up the lemon from the basket and carry it to the basket on the table behind you.
35	10	13,370	0.12	Using one hand, pick up the red apple and place it in the basket on the table behind you.
36	10	12,395	0.11	Take the water bottle from the basket and place it in the basket on the table behind you.
37	10	10,974	0.10	Using one hand, pick up the yellow plush toy from the basket and place it in the basket on the table behind you.
38	38	56,899	0.53	Go to the table and take the fruit out of the cabinet.
39	73	112,331	1.04	Open the refrigerator door, put the water bottle you are holding inside, and close the door.
40	10	8,759	0.08	Using one hand, take the banana from the current basket and place it in the basket on the table to your left.
41	10	9,593	0.09	Pick up the bread slice with one hand and place it in the basket on the table to your left.
42	10	8,738	0.08	Using one hand, lift the lemon from the basket and put it in the basket on the table to your left.
43	9	8,355	0.08	With one hand, pick up the red apple and place it in the basket on the table to your left.
44	10	9,297	0.09	Take the water bottle from the basket with one hand and place it in the basket on the table to your left.
45	10	9,316	0.09	Using one hand, take the yellow plush toy from the basket and place it in the basket on the table to your left.
46	10	9,087	0.08	Take the banana from the current basket with one hand and place it in the basket on the table to your right.
47	10	10,467	0.10	Using one hand, pick up the bread slice and place it in the basket on the table to your right.
48	10	8,989	0.08	Pick up the lemon with one hand and put it in the basket on the table to your right.
49	8	8,133	0.08	Using one hand, pick up the red apple and place it in the basket on the table to your right.
50	10	10,636	0.10	With one hand, take the water bottle from the basket and place it in the basket on the table to your right.
51	10	9,038	0.08	Pick up the yellow plush toy with one hand and place it in the basket on the table to your right.
52	20	22,520	0.21	Using both hands, lift the blue plastic basket from the current table and place it on the table behind you.
53	20	16,245	0.15	Using both hands, pick up the folded umbrella and place it on the table behind you.
54	20	15,702	0.15	Using both hands, lift the open cardboard box and place it on the table behind you.
55	5	4,517	0.04	Using both hands, lift the blue plastic basket and place it on the table to your left.
56	10	8,612	0.08	With both hands, pick up the blue storage crate and place it on the table to your left.
57	5	5,179	0.05	Using both hands, take the folded umbrella and place it on the table to your left.
58	5	4,543	0.04	Using both hands, lift the open cardboard box and set it down on the table to your left.
59	10	14,499	0.13	With both hands, pick up the sealed cardboard box and place it on the table to your left.
60	10	8,911	0.08	Using both hands, lift the white drawstring bag and place it on the table to your left.
61	5	4,372	0.04	With both hands, lift the blue plastic basket and place it on the table to your right.
62	10	7,512	0.07	Using both hands, pick up the blue storage crate and place it on the table to your right.
63	5	5,713	0.05	With both hands, take the folded umbrella and place it on the table to your right.
64	5	4,728	0.04	With both hands, lift the open cardboard box and place it on the table to your right.
65	10	15,894	0.15	Using both hands, pick up the sealed cardboard box and place it on the table to your right.
66	10	9,030	0.08	With both hands, lift the white drawstring bag and place it on the table to your right.
67	43	48,463	0.45	Unscrew the cap from the test tube.

How to load one episode

import json
from pathlib import Path

import pyarrow.parquet as pq

root = Path("MM-30")  # or the snapshot from huggingface_hub
episode_index = 0
chunk_index, file_index = divmod(episode_index, 1000)

table = pq.read_table(
    root / f"data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet"
)
state = table.column("observation.state").combine_chunks().flatten().to_numpy()
state = state.reshape(table.num_rows, 80)

videos = {
    name: root / f"videos/{name}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4"
    for name in ("primary", "wrist_left", "wrist_right")
}
instructions = json.loads((root / "meta/task_instructions.json").read_text())

Episode metadata (length, task string, global index range, video timestamps) is in meta/episodes/chunk-000/file-000.parquet.

pip install -U huggingface_hub
hf download Kivy/MM-30 --repo-type dataset --local-dir MM-30

Checked before release

• 2,443 parquet files and 2,443 videos for each of the three cameras are present.
• Every video is 640×480 H.264 yuv420p, and its frame count equals the episode length.
• Global index runs contiguously from 0 through 3,383,138.
• State and action are finite, equal to each other, and zero on every dimension marked invalid.
• Rot6d blocks are orthonormal within 1e-3. Grippers lie in [0, 1].
• Camera folders match the source cameras: primary ← head, wrist_left ← left wrist, wrist_right ← right wrist.
• Color-corrected frames match the channel-swapped source JPEG to about 3 gray levels, which is the expected H.264 gap.

Limitations

• No separate action or command channel. See the section above.
• The absolute end-effector frame was not verified with a kinematic model.
• Base-velocity axis order and lift units follow the collector's definition ([vx, vy, wz], metres) and were not re-derived from a controller log in this release.
• Camera streams are aligned by frame index. Each stream has its own capture timestamp; the residual is under one frame at 30 Hz.
• Videos were encoded twice (capture JPEG to H.264, then the color correction pass).
• tasks.parquet keeps one instruction per task. Other phrasings are only in meta/task_instructions.json.

License

CC BY 4.0. You may use, share, and adapt the dataset, including commercially, with attribution.

Citation

@misc{mm30_2026,
  title={MM-30: A Real-World Mobile Manipulation Dataset},
  author={{MM-30 Contributors}},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/Kivy/MM-30}},
  note={Real-world mobile bimanual manipulation demonstrations}
}