Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
9
108
author
stringlengths
2
27
kind
stringclasses
2 values
family
stringclasses
13 values
robot
stringclasses
81 values
version
stringclasses
3 values
episodes
int64
5
14.4k
frames
int64
50
2.82M
fps
float64
3
60
mean_episode_s
float64
0.41
250
median_episode_s
float64
0.03
223
⌀
fixed_length
stringclasses
2 values
00maker/lerobot_demo1
00maker
teleop
pick-place
so101_follower
v2.1
30
8,199
30
9.11
9.05
no
00ri/eval_act_so100_battery_bin_center
00ri
policy
pick-place
so100
v2.1
10
6,069
30
20.23
26.73
no
00ri/so100_battery_bin_center
00ri
teleop
pick-place
so100
v2.1
50
17,715
30
11.81
11.58
no
0mrrabbit0/so101_place_block_in_drawer
0mrrabbit0
teleop
open-close
so_follower
v3.0
73
48,022
30
21.93
null
null
0xNOY/eval_snvla_dn_so101_wn_aug
0xNOY
policy
pour
so101_follower
v3.0
30
22,215
30
24.68
null
null
0xNOY/eval_snvla_so101_wn_aug
0xNOY
policy
pour
so101_follower
v3.0
30
21,683
30
24.09
null
null
0xNOY/so101_wn
0xNOY
teleop
pour
so101_follower
v3.0
133
103,433
30
25.92
null
null
1006labs/so101-table-clean-mujoco
1006labs
teleop
pick-place
so101_follower
v2.1
14,413
2,043,077
30
4.73
5.67
no
1145wang/so100_test
1145wang
teleop
other
so100
v2.0
10
3,102
30
10.34
9.82
no
11dasfc/so101_grab_merged
11dasfc
teleop
pick-place
so101_follower
v3.0
70
23,860
30
11.36
null
null
123Wei123/ball-in-hole
123Wei123
teleop
insert
so_follower
v3.0
10
5,402
30
18.01
null
null
12e21/so101_pick_box
12e21
teleop
pick-place
so101_follower
v3.0
49
27,114
30
18.44
null
null
12e21/so101_pick_orange
12e21
teleop
pick-place
so101_follower
v3.0
31
20,558
30
22.11
null
null
18houston2/eval_so101_pick1
18houston2
policy
pick-place
so101_follower
v2.1
8
5,101
30
21.25
19.05
no
18houston2/eval_so101_test3
18houston2
policy
pick-place
so101_follower
v2.1
6
6,970
30
38.72
43.65
no
1909398036ljy/so100_test5
1909398036ljy
teleop
pick-place
so100
v2.1
40
16,579
30
13.82
13.65
no
1977kantomoyuki/so101_test_20260920_143445
1977kantomoyuki
teleop
pick-place
so_follower
v3.0
10
5,979
30
19.93
null
null
1lyz123576/so101_test
1lyz123576
teleop
pick-place
so101
v2.1
135
64,505
30
15.93
15.77
no
1lyz123576/so101_test-1
1lyz123576
teleop
pick-place
so101
v2.1
35
11,863
30
11.3
11
no
2006gk/rollout_so101_pick_tissue_20260608_053757
2006gk
policy
pick-place
so_follower
v3.0
8
14,363
30
59.85
null
null
33dre/so101_dataset_PAP5
33dre
teleop
pick-place
so101_follower
v3.0
30
26,301
30
29.22
null
null
356c/so100_duck_reposition_1
356c
teleop
pick-place
so100
v2.1
60
48,776
30
27.1
27.62
no
356c/so100_nut_test_1
356c
teleop
pick-place
so100
v2.1
60
31,944
30
17.75
17.73
no
356c/so100_nut_test_11
356c
teleop
pick-place
so100
v2.1
100
38,691
30
12.9
12.82
no
3b4t4/so101_pickplace_3color_20260601_135201
3b4t4
teleop
pick-place
so_follower
v3.0
150
74,729
30
16.61
null
null
4legends/bimanualSO101CutBanana
4legends
teleop
other
bi_so100_follower
v3.0
15
3,835
30
8.52
null
null
5hadytru/so101_bench_sim
5hadytru
teleop
press
so101_follower
v3.0
3,629
1,996,120
30
18.33
null
null
5hadytru/so101_bench_sim_WM_v2.1
5hadytru
teleop
pick-place
so101_follower
v2.1
3,629
1,996,120
30
18.33
13.6
no
5hadytru/so101_grasp_1
5hadytru
teleop
pick-place
so101_follower
v3.0
210
121,831
30
19.34
null
null
5hred/tutorial
5hred
teleop
pick-place
so100
v2.0
5
1,510
24
12.58
12.08
no
6666lll/so100-pusht
6666lll
teleop
push
so100_arm
v3.0
208
25,923
10
12.46
null
null
777im/so101_pick_plug_insert_v5_v6
777im
teleop
insert
so_follower
v3.0
72
53,488
30
24.76
null
null
777im/so101_pick_plug_insert_v5_v6_draft
777im
teleop
insert
so_follower
v3.0
75
55,020
30
24.45
null
null
7Iun/so101_pick_and_place_20260811_100634
7Iun
teleop
pick-place
so_follower
v3.0
5
8,914
30
59.43
null
null
7Iun/so101_pick_and_place_v2_20260814_131803
7Iun
teleop
pick-place
so_follower
v3.0
9
16,070
30
59.52
null
null
7alexzhang7/eval_so101-turn-quarter-lever-to-left-light-diffusion
7alexzhang7
policy
other
so101_follower
v3.0
20
11,340
30
18.9
null
null
7alexzhang7/so101-turn-lever-right-to-left-max-fov
7alexzhang7
teleop
other
so101_follower
v3.0
50
16,982
30
11.32
null
null
7alexzhang7/so101-turn-steve-lever-right-to-left
7alexzhang7
teleop
other
so101_follower
v3.0
75
29,364
30
13.05
null
null
7jep7/rook_to_d4_v5
7jep7
teleop
pick-place
so101
v2.1
20
10,011
30
16.68
15.78
no
7jep7/rook_to_d4_v6
7jep7
teleop
pick-place
so101
v2.1
30
8,942
30
9.94
9.97
no
A4596-GMAIL/so101_pick_toy_251212-2
A4596-GMAIL
teleop
pick-place
so101_follower
v3.0
5
4,476
30
29.84
null
null
A4596-GMAIL/so101_pick_toy_251212-3
A4596-GMAIL
teleop
pick-place
so101_follower
v3.0
5
4,472
30
29.81
null
null
A4596-GMAIL/so101_pick_toy_251215-1
A4596-GMAIL
teleop
pick-place
so101_follower
v3.0
10
7,376
30
24.59
null
null
ABDHAM/rescuehands_table
ABDHAM
teleop
handover
so101_bimanual_sim
v3.0
101
57,173
20
28.3
null
null
AEA-HF/SOARM101_merged
AEA-HF
teleop
other
so101_follower
v3.0
50
19,764
30
13.18
null
null
AHegai/so101-table-cleanup
AHegai
teleop
pick-place
so101_follower
v3.0
80
47,513
30
19.8
null
null
AJW666/staple_box_in_marked_frame_v13_clean
AJW666
teleop
pick-place
so_follower
v3.0
50
32,544
30
21.7
null
null
AJW666/staple_box_in_marked_frame_v14_clean
AJW666
teleop
pick-place
so_follower
v3.0
50
30,484
30
20.32
null
null
AK51/eval_act_so100_test5
AK51
policy
pick-place
so100
v2.1
10
2,484
25
9.94
9.96
no
AK51/so100_test8
AK51
teleop
pick-place
so100
v2.1
30
7,470
25
9.96
9.96
yes
AMRUBARD/so100_make_coffeetest6
AMRUBARD
teleop
pour
so100
v2.1
30
39,135
30
43.48
43.93
no
AS-Robotics/Pick-and-Place-dataset_so100_pickplace
AS-Robotics
teleop
pick-place
so100_follower
v2.1
130
50,293
30
12.9
12.78
no
Aaaay-0610/so101-task-1
Aaaay-0610
teleop
pick-place
so_follower
v3.0
75
35,614
30
15.83
null
null
Aaaay-0610/so101-task-1_20260927_145342
Aaaay-0610
teleop
pick-place
so_follower
v3.0
45
22,187
30
16.43
null
null
Aadhavan/so101_bio_final
Aadhavan
teleop
pick-place
so101
v2.1
50
22,326
30
14.88
16.08
no
Aaron757/so100_test7
Aaron757
teleop
pick-place
so100
v2.1
20
13,275
30
22.12
22.2
no
AaronNewman/screwdriver_task_batch3
AaronNewman
teleop
pick-place
so100
v2.0
10
3,565
30
11.88
11.9
yes
Aaronng456/SO101_picknplace2
Aaronng456
teleop
pick-place
so101_follower
v3.0
52
47,201
30
30.26
null
null
Aaronng456/SO101_picknplace4
Aaronng456
teleop
pick-place
so101_follower
v3.0
64
58,717
30
30.58
null
null
Aayush1104/Lego_Pick_and_Place
Aayush1104
teleop
pick-place
so100
v2.1
10
4,041
30
13.47
13.6
no
Aayush1104/Lego_Place_Block_On_Another
Aayush1104
teleop
pick-place
so100
v2.1
10
7,363
30
24.54
24.15
no
AbdullahRasul/so101_pick_place
AbdullahRasul
teleop
pick-place
so101_follower
v3.0
50
16,875
30
11.25
null
null
Abiam77/so101_cube_color
Abiam77
teleop
pick-place
so_follower
v3.0
50
26,849
30
17.9
null
null
Abiam77/so101_cube_color_20261108_003255
Abiam77
teleop
pick-place
so_follower
v3.0
11
6,036
30
18.29
null
null
AdenEndure/so101_640x480_50_20260909_135618
AdenEndure
teleop
pick-place
so_follower
v3.0
7
5,584
30
26.59
null
null
AdithyaRajendran/so101_grab_brain_t2
AdithyaRajendran
teleop
pick-place
so101_follower
v3.0
239
99,845
30
13.93
null
null
Agasfen/so101-lift-cube-human-60
Agasfen
teleop
pick-place
so101_follower
v3.0
60
8,494
30
4.72
null
null
Ahi-Yu/so100_0101
Ahi-Yu
teleop
other
so100_follower
v3.0
40
17,000
30
14.17
null
null
Ahi-Yu/so100_2cam
Ahi-Yu
teleop
pick-place
so100
v3.0
50
14,694
30
9.8
null
null
Ahi-Yu/so100_strawberry_grape
Ahi-Yu
teleop
pick-place
so100
v3.0
40
32,628
30
27.19
null
null
Aiden82/so101_pangyo_pickplace_cube_260115_real
Aiden82
teleop
pick-place
so101_follower
v2.1
5
1,933
30
12.89
12.5
no
Aiden82/so101_record_test_260114_2
Aiden82
teleop
pick-place
so101_follower
v2.1
5
2,295
30
15.3
14.83
no
Aikwed/pick_orange_joint_limited
Aikwed
teleop
pick-place
so101_follower
v3.0
300
350,400
30
38.93
null
null
Aipolabs/eval_so101
Aipolabs
policy
pick-place
so_follower
v3.0
10
8,682
30
28.94
null
null
Aki233/so101_experiment1_v2
Aki233
teleop
pick-place
so_follower
v3.0
10
4,490
30
14.97
null
null
AkibaGeek/so100_test
AkibaGeek
teleop
pick-place
so100
v2.1
30
8,940
30
9.93
9.93
yes
AkshayM-X/so101_pick_place_Part1_200Epi_20260719_170814
AkshayM-X
teleop
pick-place
so_follower
v3.0
200
269,217
30
44.87
null
null
AkshayM-X/so101_pick_place_Part2_200Epi_20260720_064233
AkshayM-X
teleop
pick-place
so_follower
v3.0
200
179,509
30
29.92
null
null
AkshayM-X/so101_pick_place_Part3_200Epi_redcube_20260728_20260728_112915
AkshayM-X
teleop
pick-place
so_follower
v3.0
200
239,290
30
39.88
null
null
AkshuAriga/so101_green_screwdriver_yellow_bin_2cam
AkshuAriga
teleop
pick-place
so_follower
v3.0
75
34,442
30
15.31
null
null
AkshuAriga/so101_pi05_merged
AkshuAriga
teleop
pick-place
so_follower
v3.0
122
60,432
30
16.51
null
null
AlexTranscend/so101_earplug_cup_v1
AlexTranscend
teleop
pick-place
so_follower
v3.0
15
6,533
30
14.52
null
null
AlexTranscend/so101_earplug_cup_v3
AlexTranscend
teleop
pick-place
so_follower
v3.0
50
30,636
30
20.42
null
null
AliKhoja/so101-demo-50_20260829_234547
AliKhoja
teleop
pick-place
so_follower
v3.0
50
37,819
30
25.21
null
null
Alkatt/eval_smolvla_so101_CubeToBowl_ASN_FORMAL_SEEN_POSITIONS_20251120
Alkatt
policy
pick-place
so101_follower
v3.0
130
56,003
30
14.36
null
null
Allenljh/so101_lego_data
Allenljh
teleop
pick-place
so101_follower
v2.1
5
2,259
30
15.06
15.07
yes
Alloune/SO-101_box_to_bin_20260521_204142
Alloune
teleop
pick-place
so_follower
v3.0
5
3,430
30
22.87
null
null
Alloune/SO-101_box_to_bin_20260522_090836
Alloune
teleop
pick-place
so_follower
v3.0
20
11,853
30
19.76
null
null
AlvinAi/insert_carrot_into_the_hole_hil_v2_r1
AlvinAi
teleop
insert
so_follower
v3.0
83
38,182
30
15.33
null
null
AmberHyunKIM/so101_red_blue_cube_sorting
AmberHyunKIM
teleop
sort
so101_follower
v3.0
80
43,878
30
18.28
null
null
AmberHyunKIM/so101_red_blue_v2
AmberHyunKIM
teleop
pick-place
so101_follower
v3.0
80
40,327
30
16.8
null
null
AmberHyunKIM/so101_red_only
AmberHyunKIM
teleop
pick-place
so101_follower
v3.0
60
30,106
30
16.73
null
null
AndrejOrsula/lerobot_double_ball_stacking
AndrejOrsula
teleop
stack
so100
v2.1
51
13,476
30
8.81
8.67
no
AndrejOrsula/lerobot_double_ball_stacking_random
AndrejOrsula
teleop
stack
so100
v2.1
41
14,348
30
11.67
11.43
no
AndrejOrsula/lerobot_triple_ball_stacking
AndrejOrsula
teleop
stack
so100
v2.1
101
48,652
30
16.06
16.17
no
AndrewNoviello/so101-dataset-cleaned
AndrewNoviello
teleop
other
so101_follower
v2.1
79
34,886
30
14.72
null
null
AndrewNoviello/so101-dataset-lt
AndrewNoviello
teleop
pick-place
so101_follower
v3.0
60
26,549
30
14.75
null
null
AndrewNoviello/so101-dataset-new
AndrewNoviello
teleop
pick-place
so101_follower
v3.0
79
34,886
30
14.72
null
null
Angry-Lasagna/greycube_bluebox
Angry-Lasagna
teleop
pick-place
so101
v2.1
50
14,812
30
9.87
9.68
no
AnonymousMouse404/so101_pick_in_place_20260604_133249
AnonymousMouse404
teleop
pick-place
so_follower
v3.0
39
8,484
30
7.25
null
null
End of preview. Expand in Data Studio

SO-arm task clock

How long is one recorded episode of a common task on an SO-100 or SO-101 arm? This is a dated snapshot built from the public metadata of LeRobot datasets on Hugging Face, listed on 2026-11-18.

Sample. 14,839 SO-arm datasets from 2,830 authors were in the listing (rule below). We drew up to three per author (5,872 datasets) and read their metadata. 109 could not be read; 1,984 were read and left out (fewer than 5 episodes, or no frame or fps count); 3,779 were analysed. By repository name, 3,447 of those look like teleoperated recordings (2,065 authors) and 332 look like policy rollouts.

Seconds per recorded episode, likely-teleop datasets. For each author we take the median of their datasets' mean episode times within the family, so one busy account counts once; the percentiles are across authors.

task family authors episodes p10 s p50 s p90 s episodes an hour at p50 (illustrative)
pick and place 1,758 229,441 9.0 16.7 29.9 215
stack 76 26,556 9.1 19.0 38.7 190
sort 47 10,557 10.7 23.7 40.3 152
fold (cloth, towel, shirt) 42 4,830 11.6 27.1 77.7 133
insert (peg, plug, slot) 35 4,957 11.2 15.0 26.1 241
push 31 7,712 7.1 14.9 29.9 241
press (button, switch) 23 6,427 7.9 13.1 19.9 276
pour or scoop 19 1,857 15.8 25.9 42.4 139
hand over 15 1,188 11.3 19.9 27.7 181

The last column is 3,600 divided by the p50: the rate of back-to-back recording at the median episode length, with no resets. It is an illustration, not a measured throughput, and a recorded episode need not end in a finished task.

Likely policy rollouts are counted apart. Only pick and place has enough rollout-name authors (210): median 23.4 s an episode against 16.7 s in the likely-teleop group. The records carry no success label, so this compares recorded time between two name-based groups, not time to succeed.

What you can use it for

  • Planning a recording session. In this snapshot, 1,595 of 1,758 authors' typical pick-and-place episode is under 30 s, and 50 episodes at the median are about 14 minutes of recorded episode time, before resets.
  • Choosing an episode length. The p10 to p90 range of your task family shows how long other owners' episodes ran. Recorded length can be set by the recorder's time limit (--dataset.episode_time_s in LeRobot) as well as by the task, so treat the table as a reference range, not a target.
  • Comparing a rollout. A rollout far outside the p10-p90 range of its family is worth a look at the video; the table cannot say why it ran long.
  • Choosing a task to teach. prices.tsv sets the folding row next to public prices for the work it imitates: award-derived estimates of USD 0.41 to 1.09 per pound for the whole institutional laundry service at three US Veterans Affairs medical centres (the 0.41-0.53 estimate assumes the VISN 1 schedule quantity is yearly; the file gives each calculation), and a manufacturer's rating of up to 1,200 towels an hour for automatic towel folders.

Files

load_dataset("CyberNative-AI/so-arm-task-clock") returns datasets.tsv as the train split: 3,779 rows, 12 columns, with NA in fixed_length read as null. That uses the configs entry above, which datasets 2.14 and later read (we checked 2.14.0, 3.6.0 and 5.2.0). Older versions ignore it and fail on the mixed tables; there, pass data_files="datasets.tsv" (checked on 2.13.2). Download the other files directly.

  • clock.tsv: the table above, plus the families we did not publish and why. Written by analyse.py.
  • datasets.tsv: one row per analysed dataset (3,779): id, author namespace, kind (teleop or policy, a repository-name proxy), family, robot type, LeRobot version, episodes, frames, fps, mean and median episode seconds, fixed_length. No task text, video or frame data.
  • prices.tsv: the public prices joined to the families they imitate, with the record, the calculation and a link.
  • labels.tsv: our hand labels and the rule's family for the 59-dataset rule check.
  • sample_ids.txt: the 5,872 dataset ids drawn.
  • scripts/, Python 3 standard library only:
script reads writes
list.py OUT [MAX_PAGES] Hub listing filter=LeRobot creates OUT/; appends OUT/datasets.jsonl; replaces OUT/cursor.txt (resume point)
sample.py OUT/datasets.jsonl OUT the listing replaces OUT/so_ids.txt, OUT/sample_ids.txt
fetch.py OUT/sample_ids.txt OUT/meta.jsonl each dataset's meta/ files appends OUT/meta.jsonl (skips ids already done)
analyse.py OUT/meta.jsonl RESULT [OUT/sample_ids.txt] meta.jsonl creates RESULT/; replaces clock.json, clock.tsv, datasets.tsv
labelcheck.py LABEL_META.jsonl labels.tsv fetch output for the label ids nothing
classify.py, hub.py (imported) nothing

Use a new OUT and RESULT directory for a fresh run. The scripts only send GET requests to huggingface.co and Hugging Face file storage (*.huggingface.co, *.hf.co) over HTTPS; any other destination, including a redirect or a next-page link, is refused. Redirects are followed by hub.py itself (at most 5, redirect bodies are not read), and each final response is size-limited before it is parsed. fetch.py paces itself to about 6 requests a second and backs off on a rate-limit reply. A run today reads today's Hub, so counts will differ from this snapshot. Run on the retained metadata of this snapshot, analyse.py reproduces datasets.tsv byte for byte.

How it was made, and its limits

  • SO-arm rule. A listed dataset counts as SO-arm when the case-insensitive pattern so[\W_]?10[01]|so[\W_]?arm|so[\W_]?follower matches anywhere in its full repository id (author namespace included) or in any of its tags. 63 datasets from 8 authors match only through the author namespace. The pattern also matches inside longer words. robot_type was not used to select.
  • Draw. For each author in sorted order, their sorted ids are shuffled with one random.Random(20261118) and the first three kept (sample.py).
  • Episode time is a dataset's frames divided by fps and by episodes: the mean recorded time per episode. Recording ends when the person ends the episode or when the recorder's time limit runs out, and some recorders include pauses.
  • fixed_length is yes when the usable episode lengths read for a dataset all lie within 2 frames of each other. It is computed only for v2 datasets, from the positive length values in the first 5,000 lines of meta/episodes.jsonl, and only when there are 5 or more of them (NA otherwise). Those lengths are not always every episode: in 6 datasets with a value, the number read differs from the reported episode total (one of them is a yes). It shows near-constant length; whether a time limit caused it is unknown. Among 979 likely-teleop datasets with those lengths, 183 are yes. Leaving them out moves a published family's p50 by at most 0.5 s (stack, 19.0 to 19.5 s; clock.tsv). The v3 datasets, most of the sample, are not checked.
  • Teleop or rollout is a guess from the repository name: a name with eval, rollout or infer (so also inference) at the start of the name or after a non-letter is a likely policy rollout; every other dataset is counted as likely teleop. The metadata does not record the controller, so a row can be in the wrong group.
  • Families come from fixed keyword rules (classify.py) on the dataset name and the task strings, first match wins, not from what happened in the video. Rows the rules do not match are other (166 likely-teleop authors). "Hand over" means the name or task says hand over, pass, or left, right or other hand; it does not establish two arms. Task strings come from meta/tasks.jsonl for v2 datasets; for v3 they are printable text fragments taken from meta/tasks.parquet without a parquet reader, so a string can be split or missed.
  • Rule check. 90 SO-arm ids were drawn at random (seed 20261118); the 59 that were readable with 5 or more episodes were hand-labelled from name and task strings. The final rules agree on 55 of 59 (93%); labels.tsv holds both labels and labelcheck.py prints the four misses. The metadata it was run on is not shipped (SHA-256 85ba9842928efb19e7bea1b287058a9a05498ef60785e71e8dc1776047a4c5b1); fetch.py on those ids gives today's version.
  • Purity reading. For each family with 10 or more authors we read 10 random members by hand; a family under 8 of 10 is not published. Pick and place, pour and hand over scored 10, stack and sort 9, push and press 8. Fold (4), insert (7) and open-close (3) failed; their rules were tightened once to require the action and were re-read on a fresh 10: fold 10, insert 10, open-close 7 (not published). Draw was tightened the same way and scored 8 of 9; it has 7 authors and is not published.
  • Teleop is not autonomy. These are recorded times. They say nothing about how often a person or a policy succeeded.
  • Prices cover folding only, and the laundry prices are for the whole service, not folding alone. A search of US federal notices found no public price for kitting, assembly or mail sorting by the piece.
  • Ids and removal. Every id here appeared in the public, anonymous Hub listing on 2026-11-18. A dataset made private or deleted since then still appears in this snapshot; ask and we will remove it.

Licence

Apache-2.0 covers what this project publishes here: the tables, the labels and the scripts. The datasets these rows describe belong to their authors and keep their own licences and terms on the Hub.

Prepared by CyberNative AI LLC, an AI-run company, using AI agents. Corrections, including a dataset in the wrong family: open a discussion on this dataset or write to hello@cybernative.ai. We date each correction in this card.

Downloads last month
-