UMI-Benchmark-v1
UMI-Benchmark-v1 contains 20,000 real-world robot manipulation sessions across 10 tasks:
4 single-arm tasks and 6 dual-arm tasks.
Dataset Summary
| Category |
Tasks |
Sessions |
| Single-arm |
4 |
9,988 |
| Dual-arm |
6 |
10,012 |
| Total |
10 |
20,000 |
Tasks
| ID |
Folder |
Type |
Task |
Sessions |
| T1 |
single_arm_task1 |
Single-arm |
Stack Baskets |
3,000 |
| T2 |
single_arm_task2 |
Single-arm |
Trash Bag |
1,991 |
| T3 |
single_arm_task3 |
Single-arm |
Stamp Ink |
2,996 |
| T4 |
single_arm_task4 |
Single-arm |
Remote Storage |
2,001 |
| T5 |
dual_arm_task1 |
Dual-arm |
Pour Beans |
1,600 |
| T6 |
dual_arm_task2 |
Dual-arm |
Pack & Carry |
1,600 |
| T7 |
dual_arm_task3 |
Dual-arm |
Turntable Pick |
2,012 |
| T8 |
dual_arm_task4 |
Dual-arm |
Mahjong Sort |
1,600 |
| T9 |
dual_arm_task5 |
Dual-arm |
Kitchen Rearrange |
1,600 |
| T10 |
dual_arm_task6 |
Dual-arm |
Pants Folding |
1,600 |
Scenario Counts
TASK 01 Stack Baskets / single_arm_task1
| Scenario |
Sessions |
| Blue Baskets |
985 |
| Red Baskets |
1,000 |
| Orange Baskets |
1,015 |
| Total |
3,000 |
TASK 02 Trash Bag / single_arm_task2
| Scenario |
Sessions |
| White Garbage Bag |
486 |
| Purple Garbage Bag |
497 |
| Green Garbage Bag |
500 |
| Blue Garbage Bag |
508 |
| Total |
1,991 |
TASK 03 Stamp Ink / single_arm_task3
| Scenario |
Sessions |
| Black Ink + Gray Paper |
482 |
| Black Ink + Yellow Paper |
492 |
| Black Ink + White Paper |
499 |
| Red Ink + Gray Paper |
500 |
| Red Ink + Yellow Paper |
505 |
| Red Ink + White Paper |
518 |
| Total |
2,996 |
TASK 04 Remote Storage / single_arm_task4
| Scenario |
Sessions |
| Remote B + Holder B |
489 |
| Remote A + Holder B |
500 |
| Remote B + Holder A |
501 |
| Remote A + Holder A |
511 |
| Total |
2,001 |
TASK 05 Pour Beans / dual_arm_task1
| Scenario |
Sessions |
| Red Beans Scene B |
395 |
| Yellow Beans Scene B |
400 |
| Yellow Beans Scene A |
400 |
| Red Beans Scene A |
405 |
| Total |
1,600 |
TASK 06 Pack & Carry / dual_arm_task2
| Scenario |
Sessions |
| Purple Ball Scene B |
387 |
| Green Ball Scene B |
397 |
| Green Ball Scene A |
403 |
| Purple Ball Scene A |
413 |
| Total |
1,600 |
TASK 07 Turntable Pick / dual_arm_task3
| Scenario |
Sessions |
| Yellow Can Scene B |
495 |
| Green Can Scene B |
500 |
| Green Can Scene A |
503 |
| Yellow Can Scene A |
514 |
| Total |
2,012 |
TASK 08 Mahjong Sort / dual_arm_task4
| Scenario |
Sessions |
| Mahjong Combination 2 Scene B |
389 |
| Mahjong Combination 1 Scene B |
398 |
| Mahjong Combination 1 Scene A |
402 |
| Mahjong Combination 2 Scene A |
411 |
| Total |
1,600 |
TASK 09 Kitchen Rearrange / dual_arm_task5
| Scenario |
Sessions |
| Pink Cup-Plate Set Scene B |
393 |
| Green Cup-Plate Set Scene B |
400 |
| Pink Cup-Plate Set Scene A |
400 |
| Green Cup-Plate Set Scene A |
407 |
| Total |
1,600 |
TASK 10 Pants Folding / dual_arm_task6
| Scenario |
Sessions |
| Green Pants Scene B |
391 |
| Blue Pants Scene B |
399 |
| Blue Pants Scene A |
401 |
| Green Pants Scene A |
409 |
| Total |
1,600 |
File Layout
Data is stored as tar shards under data/.
data/
single_arm_task1/
single_arm_task2/
single_arm_task3/
single_arm_task4/
dual_arm_task1/
dual_arm_task2/
dual_arm_task3/
dual_arm_task4/
dual_arm_task5/
dual_arm_task6/
Most tasks are sharded by scenario. For example:
data/dual_arm_task2/T6_Green_Ball_Scene_A/T6_Green_Ball_Scene_A_chunk_000.tar
data/dual_arm_task2/T6_Green_Ball_Scene_A/T6_Green_Ball_Scene_A_chunk_001.tar
dual_arm_task1 is the only exception in this release. It is stored as 16 sequential
session shards:
data/dual_arm_task1/session_chunk_000.tar
...
data/dual_arm_task1/session_chunk_015.tar
The scenario counts for dual_arm_task1 are listed above.
Metadata
Metadata files are lightweight index and verification files for the tar shards. They are not
the robot trajectory data itself.
metadata/<task_name>/
scene_mapping.csv
chunk_manifest.csv
all_sessions.csv
sha256.txt
sizes.txt
sizes_bytes.txt
lists/
The metadata files provide:
| File |
Description |
scene_mapping.csv |
Scenario names, session counts, and corresponding data subdirectories. |
chunk_manifest.csv |
Tar shard paths, session counts, and first/last sessions in each shard. |
all_sessions.csv |
Per-session scenario membership and relative source paths, where available. |
sha256.txt |
SHA-256 checksums for tar shards. |
sizes.txt |
Human-readable tar shard sizes. |
sizes_bytes.txt |
Tar shard sizes in bytes. |
lists/ |
Session path lists used to build each tar shard. |
For dual_arm_task1, metadata follows the sequential shard layout:
metadata/dual_arm_task1/
sha256.txt
sizes.txt
sizes_bytes.txt
Notes
Each tar shard contains complete session folders. Use the checksum files after download to
verify that shards were transferred correctly.