Embodied AI Datasets

100 Million Sets of 3D Scene & Object Data contains 60 million sets of real-space scene data, 40 million sets of simulated synthetic scene data, and 100,000 sets of high-quality 3D object assets. The real-space scene data covers a wide range of scenarios including real estate, restaurants, cultural and museum venues, commercial retail, cultural tourism, hospitality, industrial facilities, and commercial offices; the floor plan data of simulated synthetic scenes is derived from real residential homes; high-quality 3D object assets cover seven top-level categories, namely kitchen and bath, lighting, children's furniture, home appliances, furniture, home decor, and hard furnishings, and it includes interactive object models with physical parameters. This dataset can be used for the training of embodied models.

A multimodal embodied AI dataset totaling 10,000 valid hours, collected with a six-camera Ego head-mounted device, and subjected to multiple rounds of quality inspection, AI preprocessing, and fine-grained manual action segmentation and annotation. It covers 25 categories of real-world business scenarios and outputs multi-sensor data in MCAP format along with task-action labels in multiple languages including Chinese, English, Japanese, and Korean.

A large-scale, real-world teleoperation dataset for embodied AI manipulation, captured in genuine homes with the portable uDAS 1.0 system. It spans ~50,000 hours of valid footage: ~934,000 episodes (avg. 3.2 min each) covering everyday household activities, and ~19.2 million subtask segments. Each episode synchronously records six first-person video streams — head cameras at 1408 × 1280 and wrist cameras at 640 × 480, all 30 fps H.264 — plus 30 Hz proprioceptive data: bimanual end-effector 6-DoF poses, gripper opening and torso pose. Annotation is two-level (task description + subtask timeline) and three-dimensional (scene / object / action, 8 / 13 / 30 multi-tag categories), so any single dimension can serve as an independent evaluation split. Data is delivered in LeRobot v2.1 format by episode and chunk, ready to load out of the box, directly supporting pre-training, fine-tuning and cross-scenario generalization of vision-language-action (VLA) models.

This dataset contains 1,042 egocentric video segments (approximately 35 seconds each) collected across 34 locations and 6 real-world environments, including homes, offices, and retail scenarios. Powered by self-developed VSLAM system, achieving millimeter-level positioning, multi-sensor hard-triggered synchronization (≤1ms), and a high frame rate of 60fps for RGB. It includes multi-view videos, calibrations, point clouds, SLAM trajectories, and gesture recognition results in standard formats. Designed for embodied AI, spatial perception, and 3D reconstruction, it offers high precision, diverse scenarios, and out-of-the-box usability, making it ideal for training robust perception-action models.

This dataset comprises 10,000 hours of dexterous manipulation data, capturing multimodal information across tasks such as grasping, translation, and rotation. It encompasses RGB-D visual streams, joint states, and force/torque feedback. Leveraging high-precision motion capture, the data ensures physical realism and temporal consistency in hand-object interactions, providing a robust foundation for training and validating algorithms in imitation learning, reinforcement learning, and robotic control.

135,000 Hours - Ego-centric Data_D, each set contains ①raw video, ②IMU sensor data, ③Slam and depth data, ④hand keypoints, ⑤subtask annotation, covers diverse scenes and tasks including household operations, office light tasks, supermarket/retail, and industrial/semi-structured operations. This dataset can be used for tasks like embodied AI model training.

This dataset contains 20,000 hours of egocentric multimodal data collected from diverse real-world environments, including residential, retail, and office scenarios. It covers a wide range of human activities and manipulation tasks, such as meal preparation, cleaning, storage, garment care, merchandising, and object picking. Each sample includes synchronized 4K stereo video, camera calibration parameters, 76-point full-body pose annotations, and fine-grained step-by-step action sequence labels. The dataset is suitable for robot learning, manipulation policy development, and Vision-Language-Action (VLA) models.

This dataset contains 116,048 sets of 3D handpose data, each set includes hand mask image(RGB, 24-bit), depth image(16-bit), camera intrinsic parameter file(TXT), 3D keypoints file(OBJ), mesh file(OBJ), gesture type file(TXT), keypoints demo image(JPG), and mesh demo image(JPG). The data is collected indoors, with the right hand (no handheld objects), covering both first-person and third-person perspectives, multiple gesture types, finger poses, hand overall rotation poses, individuals and Kinect devices used. This dataset does not include personally identifiable facial information, with hand mask images and depth images aligned. This dataset can be used for tasks such as handpose recognition, hand 3D reconstruction, and hand keypoints detection.