Lerobot SO-101 Manipulations Dataset

This egocentric dataset contains 4,050 hours of first-person videos capturing daily activities in home environments, designed for Physical AI, robotic systems, manipulation tasks, and egocentric vision research. The videos were captured using two hardware configurations — Pico + Motion Trackers (2,321 hours, 57.3%) and Zed + Pico + Motion Trackers (1,729 hours, 42.7%). Quaternion-based orientation from onboard sensor fusion supports 3D pose estimations and egocentric tracking across first-person perspectives.

Car Accident Video Dataset contains 5,000 annotated video clips capturing 3 seconds before, during, and after traffic accidents, providing structured data for car accident detection and prediction tasks. This video dataset supports training models for traffic detection, road accident analysis, and crash recognition, offering consistent MP4 footage for developing reliable accident detection systems.

Scene Scanning Video Dataset includes 10,000 high-resolution (1920×1080, 30 FPS) MP4 videos capturing people looking around and scanning their environment in controlled settings. This video dataset contains annotated metadata labels—such as environment type, camera motion, participant count, gender, duration, and resolution—providing structured training data for computer vision, face detection, face recognition, and object detection.

Fall Detection Dataset contains 10,000 high-resolution videos of staged human fall events recorded across indoor and outdoor settings using static and moving cameras. This camera dataset supports fall detection computer vision research by providing well-structured metadata, realistic fall scenarios, and consistent 1080p video quality for training and evaluating detection systems.

Weapon Detection Dataset contains 10,000 annotated images of staged individuals with visible weapons, collected from public CCTV footage and internet sources for training weapon detection models. This surveillance dataset includes bounding box annotations and rich metadata, supporting accurate weapons detection, small object detection, and real-time security systems across CCTV and camera-based monitoring scenarios.

This fight detection dataset contains 1,000 high-quality videos of simulated physical altercations recorded in controlled environments, captured from static and moving surveillance camera views at up to 1920×1080 resolution and 30 FPS. Designed for violence detection, action recognition, and public safety systems, this surveillance dataset includes rich metadata annotations enabling accurate camera fight analysis and training violence detection models.

This droid dataset comprises 1,000 hours of multimodal recordings of cleaning, laundry folding, and dishwashing activities, combining head-mounted video with detailed robotic data streams from seven 9-axis IMU units positioned on the chest, upper arms, forearms, and wrists. The robot dataset captures complete robotic data through 3-axis accelerometer (linear acceleration), 3-axis gyroscope (angular velocity), and 3-axis magnetometer (magnetic field orientation) sensors, using onboard sensor-fusion to generate orientation quaternions rather than raw motion signals, enabling precise motion tracking and real-world robotic learning.

This Serbian road traffic dataset contains 5,000 high-resolution (≥1080p) videos and annotated frame images with JSON bounding box annotations for vehicle detection and classification tasks. Collected from bridge vantage points, it provides detailed data on traffic flows, vehicle types (cars and minivans), and traffic density, supporting AI development for traffic prediction, road safety analysis, and computer vision model training.

This smoke and fire detection dataset offers 85 high-quality RGB videos in MP4 format with JSON annotations, including frame numbers, object coordinates, and classes for fire and smoke. The dataset supports object detection, fire and smoke recognition, and computer vision tasks, enabling deep learning models for real-time monitoring, early detection, and efficient fire management under varied environmental conditions.

Real-Time Traffic Video Dataset is a traffic dataset featuring annotated video streams from traffic cameras at urban intersections and crosswalks, designed for vehicle detection, pedestrian detection, and traffic monitoring in real-time traffic scenarios with metadata on weather, time, and traffic events to support advanced traffic surveillance and traffic prediction models.