16 Embodied AI Datasets Covering Grasping, QA, Logical and Trajectory Reasoning

High‑quality multimodal interaction datasets are essential for training and evaluating embodied AI models, reducing robot data‑collection costs and improving generalisation.

Recommended embodied‑AI datasets

TongSIM‑Asset : Open‑source simulation environment released by BIGAI (2025). Contains >25,877 task scenes and 100 high‑quality 3D environments, runs at >60 FPS, provides 3,000+ interactive objects (500+ categories) and >10 agent types. Size : not specified; Download : https://go.hyper.ai/2mwQM OmniRetarget : Full‑body robot motion‑retargeting dataset from Amazon, MIT and UC Berkeley. Includes three subsets – robot‑object, robot‑terrain, robot‑object‑terrain – totaling ~4 hours of trajectories. Provides URDF, SDF and OBJ models for visualisation only. Size : 349.61 MB; Download : https://go.hyper.ai/nT7n8 InternScenes : Large‑scale indoor simulation scene dataset (Shanghai AI Lab & Shanghai Jiao‑Tong University). ~40 k diverse scenes, 1.96 M 3D objects, 15 indoor types, 288 object categories; ~20 % objects are interactive. Size : 185.91 GB; Download : https://go.hyper.ai/VljGl FoMER Bench : Embodied reasoning benchmark from Mohamed Ben Zayed AI University, Linnëa University and ANU. >1,100 samples covering 10 tasks and 8 reasoning steps, with MCQ, TF and open‑ended questions, each paired with visual observations and step‑by‑step reasoning traces. Size : 7.03 GB; Download : https://go.hyper.ai/MlwlQ DexGraspVLA : Robot grasping dataset from Psi‑Robot. 51 human‑demonstration samples achieving >90 % success on unseen objects, lighting and backgrounds. Uses a pretrained vision‑language model for high‑level planning and a diffusion‑based low‑level controller. Size : 7.29 GB; Download : https://go.hyper.ai/nrJt9 EQA (Embodied Question Answering) : Visual‑question‑answering dataset built on House3D. Agents must navigate to gather visual evidence before answering questions (e.g., "What color is the car?"). Size : 839.6 KB; Download : https://go.hyper.ai/8zLIy EgoThink : First‑person visual‑question‑answering benchmark from Tsinghua University. 700 images sampled from Ego4D, covering six core abilities across twelve dimensions; ~20 % objects are interactive. Size : 865.29 MB; Download : https://go.hyper.ai/1heWB Open X‑Embodiment : Real‑robot dataset from DeepMind (2023). Aggregates data from 22 robot types, 527 skills and 160,266 tasks in unified RLDS format. Download : https://go.hyper.ai/cP8sJ SocialMaze : Logical‑reasoning benchmark for hidden‑role inference in multi‑agent social interactions. Designed to evaluate LLMs on deception detection and multi‑turn dialogue understanding. Size : 169.48 MB; Download : https://go.hyper.ai/uCruh BC‑Z Robot Learning : Massive robot‑learning dataset from Peking University and Shanghai University of Engineering. Provides 32.28 GB of embodied navigation data with ~110 k step‑by‑step chain‑of‑thought trajectories. Size : 32.28 GB; Download : https://go.hyper.ai/nh55W ShareGPT‑4o‑Image : Image‑generation dataset containing 92,256 samples produced by GPT‑4o, split between text‑to‑image (45,717) and text‑and‑image‑to‑image (46,539) prompts. Covers diverse styles and embodied visual‑reasoning scenes. Download : https://go.hyper.ai/cW5kz RT‑1 Robot Action : Real‑world robot dataset from Google. 13 × 7‑DOF arms collected 130 k clips (111.06 GB) over 17 months, each annotated with natural‑language instructions for tasks such as picking, opening drawers and stacking. Download : https://go.hyper.ai/Dnb74 Motions Dataset : Biomimetic arm dynamic motion dataset from the Max Planck Institute. Records from a 4‑DOF pneumatic‑muscle‑driven arm over 3.5 weeks, including pressure‑targeted motions and labelled behaviour segments. Download : https://go.hyper.ai/hzeKh BridgeData V2 : Large‑scale robot‑learning collection released jointly by UC Berkeley, Stanford, DeepMind and CMU. Contains 60,096 robot trajectories across 24 environments, each paired with natural‑language instructions to promote generalisable skill learning. Download : https://go.hyper.ai/buytZ Language‑Table : Robot language‑labelled trajectory dataset with ~600 k entries. Enables robots to follow natural‑language commands and perform end‑to‑end audiovisual‑motor skills ten times more effectively than previous datasets. Download : https://go.hyper.ai/X10ie These resources collectively provide the data needed to train, evaluate and advance embodied AI models across perception, reasoning and action domains.