Data, Hardware & Evaluation
Major Datasets
Open X-Embodiment, DROID, BridgeData V2, AgiBot World, and RoboMIND compared.
Last reviewed 2026-08-09
Five open datasets anchor robot learning in 2026, and each one is a different bet on how to collect at scale. Open X-Embodiment federated existing data across 34 labs Open X-Embodiment Collaboration 2023. DROID sent one standardized rig to 13 institutions and collected for a year Khazatsky 2024. BridgeData V2 ran a multi-site academic effort on a low-cost arm Walke 2023. AgiBot World pointed a corporate fleet at the problem AgiBot Research 2025. RoboMIND put four different robots through one collection protocol Wu 2024. The bet matters as much as the count, because it decides what the data can teach and who is allowed to use it.
The comparison table
6 of 6 datasets
| License | Source | |||||||
|---|---|---|---|---|---|---|---|---|
| AgiBot World | 2025 | 1,003,672Beta release; ~43.8 TB | n/a | 2175 deployment scenarios | n/a | 1 platformAgiBot G2 | CC BY-NC-SA 4.0 | agibot-world.com |
| Open X-Embodiment (OXE) | 2023 | 1,000,000reported as 1M+; pooled from 60 datasets | n/a | 160,266527 skills | n/a | 22 platforms22 platforms pooled across 34 labs | mixed (per constituent dataset) | robotics-transformer-x.github.io |
| RoboMIND | 2024 | 107,000incl. 5k failure demonstrations | n/a | 47996 object classes | n/a | 4 platformsFranka Emika Panda, UR5e, AgileX dual-arm, humanoid, dual dexterous hands | CC BY-NC-SA 4.0 | x-humanoid-robomind.github.io |
| DROID | 2024 | 76,000 | 350 | 86 | 564 | 1 platformFranka Panda | CC BY-NC 4.0 | droid-dataset.github.io |
| BridgeData V2 | 2023 | 60,096incl. 9,731 scripted pick-and-place rollouts | n/a | 13skills | 24 | 1 platformWidowX 250 | CC BY 4.0 | rail-berkeley.github.io |
| AgiBot World 2026 | 2026 | n/a | n/a | n/a | n/a | 1 platformAgiBot G2 | CC BY-NC-SA 4.0 | huggingface.co |
Read the table with two caveats. First, episodes are not hours. A trajectory can be eight seconds or eight minutes, and only DROID publishes an hour count, so every other hours cell is n/a rather than an estimate dressed up as data. Second, the newest release, AgiBot World 2026, publishes a 13.2 TB total file size and nothing else as of August 2026: no episode count, no hours, no task breakdown AgiBot 2026. Terabytes measure storage, not supervision, so those cells stay empty until AgiBot publishes the counts.
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 tasks Open X-Embodiment Collaboration 2023. It is the broadest cross-embodiment resource available, and the RT-X experiments showed what breadth buys: RT-1-X, trained on the pool, improved success rates by 50% over robot-specific baselines on in-distribution tasks and showed transfer to embodiments it barely saw Open X-Embodiment Collaboration 2023.
The cost of federation is heterogeneity. The constituent datasets vary wildly in action spaces, control frequencies, and camera setups, and the licensing is mixed per constituent, which complicates commercial use. By 2026 a common critique holds that much of the pooled data is low quality, with no good method for quantifying data quality in imitation learning Reuss 2025. Most generalist policies train on it regardless, because nothing else is as broad (see VLA models).
DROID: depth on one platform
DROID is the largest open dataset collected on a single robot platform: 76,000 trajectories totaling 350 hours, across 564 scenes and 86 tasks, gathered by 50 operators at 13 institutions over 12 months Khazatsky 2024. Every trajectory came from the same rig, a Franka Panda with two ZED 2 cameras and a ZED Mini wrist camera, teleoperated through a Meta Quest 2. One platform means one action space and one camera geometry, which makes the data far easier to consume than a federation, at the price of embodiment lock-in.
The payoff showed up in controlled comparisons: diffusion policies trained on DROID beat policies trained on the next-best dataset by 22% in-distribution and 17% out-of-distribution Khazatsky 2024. The project also maintains the data after release: improved camera calibrations for 36,000 episodes shipped in April 2025, and language annotations for 75,000 episodes in December 2024 Khazatsky 2024. The license is CC BY-NC 4.0, which rules out commercial training runs.
BridgeData V2: the academic workhorse
BridgeData V2 is the mid-scale counterpoint: 60,096 trajectories on a WidowX 250, a 6-DoF arm cheap enough for any lab to buy Walke 2023. Of those, 50,365 are teleoperated demonstrations collected with a VR controller at 5 Hz, averaging 38 timesteps, about eight seconds per trajectory; the remaining 9,731 are scripted pick-and-place rollouts. The data spans 24 environments, mostly toy kitchens, and 13 skills from pick-and-place to cloth folding Walke 2023.
Two things keep BridgeData V2 on every shortlist. The license is CC BY 4.0, the most permissive in this table. And the paper shipped trained baselines, from goal-conditioned behavior cloning to ACT and an RT-1 fine-tune, which made the dataset a default benchmark pool for academic generalists Walke 2023. Most of it also ships inside OXE as one of the 60 constituents, so policies trained on the pool have already seen it Open X-Embodiment Collaboration 2023.
AgiBot World: the corporate fleet
AgiBot World is what a robot manufacturer can collect when the collection fleet is the product line: the Beta release holds 1,003,672 trajectories across about 43.8 TB, all from the AgiBot G2 platform with a multi-camera rig AgiBot Research 2025. The paper reports over a million trajectories across 217 tasks in five deployment scenarios, and claims that policies pretrained on AgiBot World improve 30% on average over policies pretrained on OXE, in-domain and out-of-distribution AgiBot Research 2025. That figure is vendor-reported; no independent replication exists. The companion GO-1 model needs about 7 GB of VRAM for inference and 70 GB for a full fine-tune, which makes the dataset the only realistic way to study the model AgiBot Research 2025.
The catch is the license and the opacity. Everything ships under CC BY-NC-SA 4.0, so commercial use is out. And the hour count is unpublished, as it is for OXE; the hundred-thousand-hour figure that circulates is an estimate, not a measurement. AgiBot World 2026, the newest release, is 13.2 TB collected entirely in real-world scenes, but publishes no episode, hour, or task counts at all as of August 2026 AgiBot 2026.
RoboMIND: multi-embodiment under one protocol
RoboMIND splits the difference between OXE's federation and DROID's monoculture: 107,000 trajectories across 479 tasks and 96 object classes, collected on four different robots, a Franka Emika Panda, a UR5e, an AgileX dual-arm, and a humanoid with dual dexterous hands, but all under one collection platform and one standardized protocol Wu 2024. Cross-embodiment data without the cross-lab inconsistency.
Two details set it apart. It includes 5,000 real-world failure demonstrations with annotated causes, and failure data is nearly absent from every other dataset in this table, which are collections of successes. And it ships a digital twin of the tasks and assets in Isaac Sim, so additional data can be collected and policies evaluated in simulation against the same distribution Wu 2024. The license is CC BY-NC-SA 4.0.
How to read dataset claims
Convert to hours wherever you can. DROID's 350 hours across 76,000 trajectories works out to about 17 seconds each; BridgeData V2's 38 steps at 5 Hz is about 8 seconds. Episode counts flatter datasets of short trajectories, and a million-episode claim can hide less total interaction than a hundred-thousand-episode dataset of long-horizon tasks. The arithmetic behind this is in the data bottleneck.
Check the license before the leaderboard. BridgeData V2's CC BY 4.0 permits commercial training; DROID, AgiBot World, and RoboMIND all carry non-commercial terms. A policy trained on NC data is an NC policy in practice, whatever the license lawyers eventually decide.
Finally, weigh breadth against consistency. OXE's federation maximizes diversity and inherits every constituent's quirks; DROID's monoculture gives up embodiment coverage for a clean action space. The scaling-laws work says environment and task diversity dominates raw volume, and that multi-embodiment pretraining may be optional when task diversity is high Shi 2025. If that result holds, the DROID and RoboMIND bets age better than the OXE bet. The field has not settled it.