Evaluation & benchmarks
17Held-out tasks and benchmarks for measuring policy behavior — not for training.
Research topic · VLA training data
Data for vision-language-action models.
A VLA pipeline needs different data at different stages. This topic separates pretraining corpora, teleoperated demonstrations, action-annotated human video, post-training traces, and evaluation sets — so a dataset's stage is a decision, not an assumption.
75 records with action-side signals68 open accessStage classification reviewed
Which stage is this dataset for?
Inclusion needs one thing a VLA policy cannot exist without: an action-side signal — robot actions, joint states, end-effector poses, proprioception, or teleoperation traces. Stage labels below are selection aids from each record's own text; records that do not state a stage stay under “Check the record”.
01
Pretraining
Broad multi-task and cross-embodiment corpora. Breadth and embodiment coverage matter more than any single task's depth.
02
Teleoperated demonstrations
Task-specific robot trajectories for imitation learning. Check action space, control rate, and camera placement against your policy.
03
Action-annotated human video
Human video with hand or action labels can augment robot data, but is not robot control data. Covered in the Ego topic, not duplicated here.
Ego data topic04
Post-training & correction
Failures, corrections, and recovery traces used after a base policy exists. Small but decisive for reliability.
05
Evaluation
Held-out tasks and benchmarks measure behavior. Never trained on — leakage here invalidates every number you report.
For how to weigh these stages against a concrete policy and embodiment, read the dataset selection guide for VLA training. Task-level browsing lives in the robot manipulation topic.
Grouped by training stage
Source-linked index
Every record links its official source for access and terms; action-side signals are listed per record so you can verify the inclusion basis yourself. A record placed in one stage today can legitimately serve another — the label is where the evidence points, not a restriction.
Pretraining corpora
9Broad multi-task or cross-embodiment corpora for pretraining before task-specific data.
Teleoperated demonstrations
19Task demonstrations teleoperated or scripted on robots, typically for imitation learning.
Post-training & correction
2Failure, correction, and refinement data used after an initial policy exists.
Check the record
28Records with action-side signals whose own labels do not state a stage yet.
Stage labels are derived from each record's own task labels and description. A stage label is selection guidance, not a usage restriction.
8 shown · 75 total
- abc_130k_v3_smokeAccess not verified
lerobot · 2026
This dataset was created using LeRobot. Dataset Structure meta/info.json: { "codebaseversion": "v3.0", "fps": 30, "features": { "observation.state": { "dtype": "float32", "shape": [ 14 ], "names": [ "leftarmjoint1", "leftarmjoint2", "leftarmjoint3", "leftarmjoint4", "leftarmjoint5"… See the full description on the dataset page:…
- Catalog-reported signals
- Robot state / trajectory · Video / RGB
- Reported license
- apache-2.0
- abc_130k_v3_trainAccess not verified
lerobot · 2026
This dataset was created using LeRobot. Dataset Structure meta/info.json: { "codebaseversion": "v3.0", "fps": 30, "features": { "observation.state": { "dtype": "float32", "shape": [ 14 ], "names": [ "leftarmjoint1", "leftarmjoint2", "leftarmjoint3", "leftarmjoint4", "leftarmjoint5"… See the full description on the dataset page:…
- Catalog-reported signals
- Robot state / trajectory · Video / RGB
- Reported license
- apache-2.0
- abc_130k_v3_valAccess not verified
lerobot · 2026
This dataset was created using LeRobot. Dataset Structure meta/info.json: { "codebaseversion": "v3.0", "fps": 30, "features": { "observation.state": { "dtype": "float32", "shape": [ 14 ], "names": [ "leftarmjoint1", "leftarmjoint2", "leftarmjoint3", "leftarmjoint4", "leftarmjoint5"… See the full description on the dataset page:…
- Catalog-reported signals
- Robot state / trajectory · Video / RGB
- Reported license
- apache-2.0
- ABC-130kOpen
UC Berkeley / MIT / Amazon FAR / CMU / XDOF · 2026
ABC-130k is an open bimanual robot manipulation dataset released with the paper 'Scalable Behavior Cloning with Open Data, Training, and Evaluation'. It contains roughly 130,919 episodes (43,090 annotated) totaling about 3,598 hours collected on bimanual YAM stations with two 6-DoF arms and parallel-jaw grippers, spanning ~195 tasks such as pick-and-place,…
130.9K episodes · parallel_jaw · aloha
- Catalog-reported signals
- Video / RGB · Language / text · Pose / calibration
- Reported license
- Apache-2.0
AgiBot (Zhiyuan Robotics) · 2026
AgiBot World 2026 is the latest release of the AgiBot World embodied-AI dataset, succeeding the earlier Alpha/Beta releases. It is collected in real-world commercial, home, and general-purpose environments using the AGIBOT G2 wheeled humanoid with dual 7-DOF force-controlled arms, capturing synchronized RGB(D), LiDAR point clouds, tactile, IMU, and…
2,976.4 hours · humanoid_other
- Catalog-reported signals
- Video / RGB · Depth / 3D · Language / text · Pose / calibration · Tactile / force
- Reported license
- custom
- AgiBotWorld2026Open
agibot-world · 2026
AgiBot World 2026 Real-World Embodied Intelligence Dataset Overview As robotics research advances into real-world scenarios, the demand for authentic, high-quality data has become increasingly urgent. Following AGIBOT WORLD's "ImageNet moment," we now release the AGIBOT WORLD 2026 dataset. Built upon massive real-world scenes, it systematically spans…
13.56TB repository size
- Catalog-reported signals
- Actions · Robot state / trajectory · Video / RGB · Language / text · Task / outcome · Simulation / future
- Reported license
- cc-by-nc-sa-4.0
AIRoA (AI Robot Association) · 2025
The AIRoA MoMa Dataset is a large-scale, real-world dataset for mobile manipulation collected with the Toyota Human Support Robot (HSR) via leader-follower teleoperation in household environments. It contains 25,469 episodes (~94 hours) sampled at 30 Hz, with synchronized RGB images from head and wrist cameras, joint states, six-axis wrist force-torque…
25.5K episodes
- Catalog-reported signals
- Video / RGB · Language / text · Tactile / force
- Reported license
- custom
CMU AirLab · 2020
ALFA (AirLab Failure and Anomaly Dataset) from CMU's AirLab is a real-world flight dataset for fault detection and isolation (FDI) and anomaly detection (AD) on a fixed-wing UAV. It contains processed data for 47 autonomous flights covering eight fault types: 23 sudden full engine (power loss) failures and 24 scenarios of seven control-surface (actuator)…
47 episodes · drone
- Catalog-reported signals
- Not recorded
- Reported license
- CC-BY-4.0
Maintain the topic
Suggest a dataset, correct a stage label, request data, or discuss sourcing, custom collection, and synthetic data with a human.
Submit a dataset
Maintainers: add a corpus or update access, license, and scale facts at the official source.
contact@openbot.aiRequest specific data
Tell us the task, embodiment, and signals you need; we point you at published records or gaps.
contact@openbot.aiDiscuss a data partnership
For data sourcing, custom collection, and synthetic data inquiries from teams and enterprises.
contact@openbot.ai