Major robot learning datasets are the open collections of recorded robot trajectories that generalist policies train on. Open X-Embodiment pools over a million trajectories from 22 embodimentsSource 1, AgiBot World holds 1,001,552 (2,976 hours) on one robot platformSource 5, and DROID holds 76,000 (350 hours) across 564 scenesSource 2.

An infographic framed by robot dataset thumbnails reading 1M episodes from 311 scenes, 34 research labs across 21 institutions, 22 embodiments, 527 skills and 60 datasets.
Open X-Embodiment pooled robot data into one training set: 1M episodes from 60 datasets, 22 embodiments and 527 skills, contributed by 34 labs at 21 institutions.
Paper figure: Open X-Embodiment Collaboration / arXiv. Licence: CC BY 4.0.
1M+
largest pool
trajectories, Open X-Embodiment
350
one platform
hours, DROID's 12-month effort
24
environments
BridgeData V2 kitchens and tabletops
4
embodiments
RoboMIND, one collection protocol

The comparison table

Filter by size
Filter by embodiment
Filter by task count
6 of 6 datasets
6 open robot-manipulation datasets. Figures the source does not publish are marked not disclosed and always sort last, in both directions; estimates appear in the prose, never in cells. Every row links to its dataset site or paper.
LicenseSource
AgiBot World20251,001,552Beta release; 1,003,672 after post-release additions2,976.42175 deployment scenariosnot disclosed1 platformAgiBot G1CC BY-NC-SA 4.0agibot-world.com
Open X-Embodiment (OXE)20231,000,000reported as 1M+; pooled from 60 datasetsnot disclosed160,266527 skillsnot disclosed22 platforms22 platforms pooled across 34 labsmixed (per constituent dataset)robotics-transformer-x.github.io
RoboMIND2024107,000107k successful, real + simulated (paper v3); card versions 1.1/1.2; 5k additional real-world failures305.5Paper v3: real + simulated; real-only and failure-set durations not separately reported47996 object classesnot disclosed4 platformsFranka Emika Panda, UR5e, AgileX dual-arm, humanoid, dual dexterous handsnot disclosedx-humanoid-robomind.github.io
DROID202476,000350865641 platformFranka PandaCC BY 4.0droid-dataset.github.io
BridgeData V2202360,096incl. 9,731 scripted pick-and-place rolloutsnot disclosed13skills241 platformWidowX 250CC BY 4.0rail-berkeley.github.io
AgiBot World 20262026not disclosednot disclosednot disclosednot disclosed1 platformAgiBot G2CC BY-NC-SA 4.0huggingface.co

Set the size filter to 1M+ episodes and one row remains; sort by Hours to compare published durations. Episodes and hours measure different things, because a trajectory can last eight seconds or eight minutes. RoboMIND reports 305.5 hours for its 107k-trajectory cohort, which mixes real and simulated dataSource 7. NVIDIA's open Physical AI dataset reports both, 15 terabytes and more than 320,000 trajectoriesSource 10.

Five robot datasets
Run counts and hours rank these datasets differently
hours recorded3001,0003,00010,000100,0001 millionruns recorded1 minute a runFineARTDROIDGalaxeaRoboMINDAgiBot World
AgiBot World has about 25 times FineART's runs but under twice its hours. FineART's runs average 2.5 minutes, the others under 20 seconds.
How this was made
Each point is one dataset's own totals: the runs it recorded (the papers say trajectories or episodes) and the hours they add up to. Both rulers step by ten times. The run lengths are hours times 3,600 divided by runs; the dashed line is where an average run lasts one minute.
  • FineART: 40,543 runs, 1,718 hours, about 2.5 minutes a run.
  • DROID: 76k runs as printed, 350 hours, about 17 seconds a run.
  • Galaxea: 100K runs as printed, 500 hours, about 18 seconds a run.
  • RoboMIND: 107k runs as printed, 305.5 hours, about 10 seconds a run.
  • AgiBot World: 1,001,552 runs, 2,976.4 hours, about 11 seconds a run.
RoboMIND's total mixes real and simulated runs, and AgiBot World is its Beta release. BridgeData V2 is left off: it reports 60,096 trajectories but no total hours, and its 5 Hz recording rate is stated for the teleoperated runs only, so any hours figure would be a guess.

Five robot datasets by runs and hours: FineART 40,543 runs and 1,718 hours, DROID 76,000 runs and 350 hours, Galaxea 100,000 runs and 500 hours, RoboMIND 107,000 runs and 305.5 hours, AgiBot World 1,001,552 runs and 2,976.4 hours. AgiBot World has about 25 times FineART's runs but 1.7 times its hours, because FineART's runs average 2.5 minutes and the other four average 10 to 18 seconds.

Runs, hours and average run length per dataset
datasetrunshoursseconds per run
FineART40,5431,718153
DROID76,00035017
Galaxea100,00050018
RoboMIND107,000305.510
AgiBot World1,001,5522,976.411
Choghari et al. 2026, Khazatsky et al. 2024, Jiang et al. 2025, Wu et al. 2024, AgiBot-World-Contributors et al. 2025.

Open X-Embodiment: breadth by federation

OXE pooled 60 existing datasets from 34 labs into one standardized RLDS format: over a million trajectories spanning 22 robot embodiments, 527 skills and 160,266 tasksSource 1. RLDS is a standard, lossless episode format for sharing and reusing datasetsSource 11.

RT-1-X trained on the pool beat each small-scale domain's original method by 50% on average, but underperformed RT-1 trained on a large-scale domain's own data, which the authors attribute to underfitting. The higher-capacity RT-2-X roughly tripled performance on out-of-distribution skills over the same model trained without the poolSource 1. Generalist policies train on it (VLA models): Octo used 800k OXE trajectoriesSource 12, and OpenVLA used 970k real-world demonstrationsSource 13.

Later cross-embodiment pools are larger or broader. ARIO reports about 3 million episodes across 321,064 tasksSource 14, and CrossFormer trained one policy on 900K trajectories from 20 embodiments, including wheeled robots, quadcopters and quadrupedsSource 15.

Federation costs consistency: constituents differ in action spaces, control rates, cameras and licenses. Moritz Reuss's review of VLA papers at ICLR 2026 calls OXE mostly low-quality data and notes that imitation learning still lacks good ways to quantify data qualitySource 8. Re-Mix's learned domain weights beat uniform weighting by 38% on average on the data used to train the RT-X modelsSource 16.

DROID: depth on one platform

DROID's distinction is scene and site diversity: 76,000 trajectories totaling 350 hours, across 564 scenes and 86 tasks, gathered by 50 operators at 13 institutions over 12 monthsSource 2. Every trajectory came from one rig, a Franka Panda with two ZED 2 cameras and a ZED Mini wrist camera, teleoperated through a Meta Quest 2. One platform gives one action space and one camera geometry, at the price of embodiment lock-in.

Single-platform collections can be larger. RT-1's fleet of 13 robots gathered about 130k episodes over 17 monthsSource 17, and Language-Table released nearly 600,000 language-labelled trajectoriesSource 18. The Galaxea Open-World Dataset holds 500 hours across 50 scenes, all from one consistent robot embodimentSource 19.

Diffusion policies trained on DROID beat policies trained on the next-best dataset by 22% in-distribution and 17% out-of-distributionSource 2. The project added language annotations for 75,000 episodes in December 2024 and improved camera calibrations for 36,000 in April 2025Source 20. The paper releases DROID under CC BY 4.0Source 2, whose deed allows commercial sharing and adaptation subject to attribution, a license link, change notices and the other license terms, and warns that other rights may limit a particular useSource 3.

BridgeData V2: the academic workhorse

BridgeData V2 holds 60,096 trajectories on a WidowX 250, a low-cost 6-DoF arm: 50,365 teleoperated demonstrations collected with a VR controller at 5 Hz and 9,731 scripted pick-and-place rollouts, across 24 environments and 13 skillsSource 4. Trajectories average 38 timestepsSource 21. Its 2021 predecessor had 7,200 demonstrations of 71 tasks in 10 environments, and joint training with 50 demonstrations of a new task in a new domain doubled success on averageSource 22.

Seven labelled camera frames from BridgeData V2 demonstrations in a toy kitchen sink, such as Put Carrot on Plate and Put Eggplant into Pot, each with a WidowX robot's black gripper above the toy objects.
Seven BridgeData V2 tasks in one toy-sink scene, such as putting a carrot on a plate, each recorded on a WidowX 250.
Paper figure: Moo Jin Kim et al. / arXiv. Licence: CC BY 4.0.

BridgeData V2's official project page states CC BY 4.0 for all dataSource 21. The paper shipped baselines from goal-conditioned behavior cloning to ACT and RT-1Source 4, and most of the data also ships inside OXESource 1.

AgiBot World: the corporate fleet

AgiBot World's Beta release holds 1,001,552 trajectories totaling 2,976 hours, all from the AgiBot G1 platform with a multi-camera rig. AgiBot reports 217 tasks in five deployment scenarios and a 30% average gain over OXE pretraining for policies pretrained on its data, a vendor figure with no independent replicationSource 5. Its repository now lists 1,003,672 trajectories and places all data and code under CC BY-NC-SA 4.0Source 23, whose NonCommercial term bars commercial useSource 24.

At about 11 seconds per trajectory, the million-trajectory headline covers less interaction than Ego4D's 3,670 hours of passive videoSource 25. AgiBot World 2026 was collected on the AgiBot G2 in real commercial and home spaces, as long-horizon restocking episodes with three annotation layers and a digital-twin simulation counterpart. Its card lists a 13.6 TB total file size but no episode or hour totalsSource 6.

RoboMIND: multi-embodiment under one protocol

RoboMIND's paper reports 107,000 successful trajectories totaling 305.5 interaction hours across 479 tasks and 96 object classes. The cohort mixes real-world and simulated demonstrations for four embodiments, a Franka Emika Panda, a UR5e, an AgileX dual-arm robot and a humanoid with dual dexterous hands, under one collection platform and standardized protocol. It also releases 5,000 real-world failure demonstrations with annotated causes and a digital twin of its tasks in Isaac SimSource 7.

RoboMIND 2.0 (December 2025) reports over 310K dual-arm trajectories from six embodiments and 739 tasks, including 12K tactile-enhanced episodes and 20K mobile manipulation trajectoriesSource 26. The public dataset card shows an Apache-2.0 tag but makes users accept access conditions before downloading, so check the terms before commercial useSource 27.

How to read dataset claims

Hours compare better than episodes. DROID averages about 17 seconds per trajectory and AgiBot World Beta about 11, so a large episode count can hide less interaction than a smaller set of long tasks (the data bottleneck). The 2025 BEHAVIOR Challenge supplies 10,000 teleoperated demonstrations totaling 1,200+ hoursSource 28, and the synthetic InternData-A1 holds over 630k trajectories and 7,433 hoursSource 29.

Modalities differ as much as size. RH20T records visual, force, audio and action data in over 110,000 contact-rich sequencesSource 30. Humanoid Everyday adds depth, LiDAR and tactile streams to 10.3k humanoid trajectories across 260 tasks, with a cloud evaluation platformSource 31. Scaling work finds task diversity worth more than per-task demonstration count, and multi-embodiment pretraining optional given high-quality single-embodiment dataSource 9.

Releases since 2025

RoboCOIN offers over 180,000 bimanual demonstrations from 15 platforms and 421 tasksSource 32. Scale AI and Hugging Face's FineART (September 2026) has 1,718 hours of bimanual teleoperation in 40,543 episodes, with 533,913 human-labelled subtasks across 151 tasksSource 38. SmolVLA trained on fewer than 30k episodes, all from community-contributed datasetsSource 33. The community-built AXIS holds 50K+ trajectories over 207 tasks, and continual pretraining on it raised π0.5's overall success rate by 5.8%Source 34.

In simulation, RoboCasa365 offers 365 tasks across 2,500 kitchens, with over 600 hours of human and over 1,600 hours of synthetic demonstrationsSource 35.

From dataset headline to training contract

A useful dataset is legal for the intended deployment, consistent in actions, units, frames and rates, complete for the claimed tasks, scenes and embodiments, and downloadable as versioned episodes traceable to a release. Datasheets for Datasets asks every dataset to document its motivation, composition, collection process and recommended usesSource 36. LeRobot's format moved to multi-episode files in 2025, after one file per episode hit file-system limits at millions of episodesSource 37.

Mixing datasets needs a canonical observation and action schema, one adapter per source and source labels kept through training, so positive transfer can be told apart from one corpus dominatingSource 1 Source 12 Source 16. The robot-learning stack supplies the versioned artifacts, and the robot-learning roadmap places dataset selection in a broader program.