Video visual-spatial VQA benchmark
Video question-answering dataset for testing visual-spatial understanding in multimodal models.
- Visual question answering
- English
Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,788 in the catalog so far, with new ones added every hour.
Answered by Gemini 4 · the chat is kept only in this tab
Video question-answering dataset for testing visual-spatial understanding in multimodal models.
Nearly 6 thousand hours of Full HD digital human video in 3D scenes for physical AI training.
Compact dataset of 206 episodes with video and time series for training robotic skills.
Video dataset tied to keystroke logs, text descriptions, and sparse 3D models.
Trimmed clips from EPIC-KITCHENS-100 with text descriptions, verbs, and objects.
Minecraft gameplay episode records with precomputed MineCLIP embeddings for fine-tuning STEVE-1-style agents.
10k to 100k raw GoPro video recordings for training robots on four household tasks.
Repackaged OpenVid-1M dataset into .tar shards for fast streaming during video model training.
An archive of episodes, logs, and video recordings of AI agent actions in Minecraft for training multimodal control models.
49K English and Chinese questions about cats and dogs with images, audio, video, and text.
Synchronized video, audio, photo, and sensor data for human action recognition.
Multi-camera face video recordings and text metadata for contactless photoplethysmography (rPPG) and health metrics assessment.
English-language test set and annotations for evaluating neural networks on long video analysis.
Large-scale extraction of texts, images, audio, and video from open datasets in a unified parquet format.
10K to 100K English questions on long videos with a hard subset.
A 50-hour egocentric video sample for video model classification and evaluation.
A complete snapshot of the Mixamo library: 2,317 animations and 114 rigged 3D characters with a unified skeleton for motion generation.
First part of a massive video corpus for robotic manipulator training by AIRoA for CoRL 2026.
25,124 English MP4 video clips for text-to-moment retrieval tasks.
A robotics dataset containing video, tabular, and time-series data created via LeRobot.
A robotics dataset containing video and tabular time-series data of the folding process.
Unitree G1 robot teleoperation recordings in the MuJoCo simulator for training robotics models.
Over 3M visual-tactile image pairs and 10k trajectories across 50 tasks for training contact-rich robotic manipulation.
Teleoperation recordings of the ALOHA robotic manipulator in simulation for robotics learning.