omert27/calibra-robot-dataset-quality-benchmark

Calibra Robot Dataset Quality Benchmark

Automated quality audits of 30 public LeRobot datasets, produced by Calibra.

Mean health score: 85.8 / 100 across 30 datasets (206–800 episodes each).

Key finding: every audited dataset has at least one critical quality flag, confirming that silent data issues are the norm, not the exception, in public robot demonstration corpora.

---

Calibra Community Dataset Quality Benchmark

Generated 2026-07-27 19:07 UTC by Calibra

Audited 30 public LeRobot datasets. 1 datasets failed to load.

Leaderboard

#	Dataset	Score	Grade	Certification	Episodes	Frames	Critical
1	aloha_mobile_elevator	91.9	A	✗ Not Certified	20	45,000	1
2	aloha_mobile_wipe_wine	91.9	A	✗ Not Certified	50	65,000	1
3	aloha_mobile_shrimp	91.3	A	✗ Not Certified	18	67,500	1
4	aloha_static_coffee	90.3	A	✗ Not Certified	50	55,000	1
5	aloha_mobile_chair	89.7	B	✗ Not Certified	55	110,000	1
6	aloha_static_candy	89.3	B	✗ Not Certified	50	35,000	1
7	aloha_mobile_wash_pan	89.3	B	✗ Not Certified	50	55,000	1
8	aloha_static_screw_driver	88.9	B	✗ Not Certified	50	20,000	1
9	aloha_static_towel	88.7	B	✗ Not Certified	50	25,000	1
10	aloha_mobile_cabinet	88.7	B	✗ Not Certified	85	127,500	1
11	aloha_static_ziploc_slide	88.4	B	✗ Not Certified	56	16,800	2
12	aloha_static_pingpong_test	88.3	B	✗ Not Certified	10	6,000	1
13	aloha_static_battery	87.9	B	✗ Not Certified	49	29,400	2
14	aloha_sim_insertion_human	87.3	B	✗ Not Certified	50	25,000	2
15	aloha_static_fork_pick_up	87.3	B	✗ Not Certified	100	60,000	2
16	aloha_static_tape	86.6	B	✗ Not Certified	50	35,000	2
17	aloha_static_thread_velcro	86.6	B	✗ Not Certified	49	34,300	2
18	aloha_static_cups_open	86.2	B	✗ Not Certified	50	20,000	2
19	aloha_sim_transfer_cube_human	86.0	B	✗ Not Certified	50	20,000	1
20	xarm_push_medium_replay	85.3	B	✗ Not Certified	800	20,000	2
21	aloha_sim_insertion_scripted	84.7	B	✗ Not Certified	50	20,000	2
22	xarm_lift_medium	83.9	B	✗ Not Certified	800	20,000	3
23	aloha_sim_transfer_cube_scripted	83.4	B	✗ Not Certified	50	20,000	2
24	xarm_lift_medium_replay	82.6	B	✗ Not Certified	800	20,000	3
25	xarm_push_medium	81.2	B	✗ Not Certified	800	20,000	1
26	unitreeh1_two_robot_greeting	80.8	B	✗ Not Certified	30	3,750	2
27	unitreeh1_rearrange_objects	79.1	C	✗ Not Certified	30	7,150	3
28	pusht	76.7	C	✗ Not Certified	206	25,650	4
29	unitreeh1_warehouse	76.1	C	✗ Not Certified	24	11,275	3
30	unitreeh1_fold_clothes	75.0	C	✗ Not Certified	38	19,000	3

Failed to Load

Dataset	Error
xarm_lift_medium_unlabeled	Dataset 'lerobot/xarm_lift_medium_unlabeled' doesn't exist on the Hub or cannot

Score Interpretation

Score	Grade	Certification
90–100	A	✓ Certified
75–89	B	✓ Certified
60–74	C	~ Provisional
40–59	D	✗ Not Certified
0–39	F	✗ Not Certified

Reproduce

pip install 'calibra-robotics[lerobot]'
python scripts/audit_lerobot_community.py

---

Files in this repository

File	Description
manifest.json	Machine-readable summary of all 30 audits
community_stats.json	Aggregate statistics for percentile comparison in the Space
leaderboard.md	This leaderboard table
lerobot/<dataset>/latest.json	Full CalibraReport JSON per dataset

Reproduce

pip install 'calibra-robotics[lerobot]'
git clone https://github.com/omertt27/Calibra
python scripts/audit_lerobot_community.py

About

Calibra is an open-source dataset quality tooling library for robotics imitation learning. It checks timing integrity, control smoothness, trajectory diversity, and more — then produces an actionable 0–100 health score with per-episode verdicts.

pip install 'calibra-robotics[lerobot]'
calibra audit hf://lerobot/pusht