Real-world data for AgiBot World 2026 robots
Corpus of real scenes with images and video for training and evaluating embodied AI and robotics.
- Robotics
- 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
Corpus of real scenes with images and video for training and evaluating embodied AI and robotics.
From 1 to 10 thousand examples for evaluating robotics VLA models in simulation.
A mix of expert RL demonstrations, image-caption pairs, and text for training a multimodal agent.
Large archive of Rise of the Tomb Raider gameplay recordings with event logs and parquet files for RL.
1,000 to 10,000 movable and rigid 3D objects with physics and joints for robot training.
Image and text mixture for the midtraining phase of the open MolmoAct robotics model by Allen Institute for AI.
Hybrid synthetic-real trajectory dataset for training robotic manipulators.
Real-world robot failure logs with visual symbols for error diagnosis and correction.
Synthetic images and texts for point tracking, 3D reconstruction, and robotics tasks.
Four LIBERO datasets with Panda robot episodes, merged and converted to the LeRobot format.
2,554 episodes in AI2-THOR and ProcTHOR simulators for evaluating the memory of embodied agents.
1,000 simulations to verify how accurately and safely robots perform household tasks.
Converted trajectory subsets from the OpenFly dataset for robotics applications.
96 single-camera driving videos and over 105 thousand frames with depth and semantic segmentation annotations.
Large multi-domain image-text corpus for 4D world modeling and robotics.
A catalog of simulation-ready 3D warehouse objects and their preview images for robotics and computer vision.
Image and timeseries records from the Allen Institute for AI for training robotics models.
An English dataset of 10K–100K examples featuring scene descriptions, trajectories, and VQA tasks for embodied AI.
Massive set of images and actions from a trained RL agent in procedurally generated Procgen benchmark games.
Over 50 hours of hand-work video from head and wrist cameras, accompanied by 3D hand poses and 122 pressure sensor readings per hand.
From 100k to 1M operational steps of top LLM agents (Claude, GPT, Gemini) in browser tasks.
A set of 1–10k English web agent trajectories for evaluating automated evaluator quality.
Processed DROID robotics dataset with optical and 3D scene flow annotations for training robot control models.
Over 100,000 urban scenes in OpenUSD format with textures for robot training.