TartanAir

Photorealistic AirSim/Unreal Engine simulation dataset for visual SLAM and navigation, with synchronized stereo RGB, depth, semantic segmentation, optical flow, camera poses, simulated LiDAR and IMU across 30 diverse environments (1037 sequences, >1M frames).

TartanAir is a large-scale synthetic dataset for visual SLAM and robot navigation, released by CMU's AirLab (Robotics Institute) and presented at IROS 2020. Data is collected in 30 photo-realistic Unreal Engine environments via the AirSim plugin, spanning urban, rural, nature, domestic, public and sci-fi scenes, with challenging conditions including day-night/lighting changes, weather (rain, snow, fog, wind), seasonal variation, moving/dynamic objects, and aggressive diverse ego-motion. It comprises 1037 long motion sequences (each 500-4000 frames) totaling over one million frames, organized hierarchically by Environment -> Difficulty (Easy/Hard) -> Trajectory (P000, P001, ...) -> modality subfolders. Each frame provides multi-modal sensor data and precise ground truth: stereo RGB (left/right, ~640x480), depth maps, semantic segmentation, optical flow (with occlusion masks), 6-DoF camera poses, simulated multi-line LiDAR point clouds, and simulated IMU. Modalities are stored in open formats: RGB as PNG, depth/segmentation/optical flow as NumPy .npy arrays, and poses as .txt files. The full dataset is up to ~3TB and is distributed via the Microsoft Azure Open Datasets platform, an AirLab Ceph/S3 endpoint, and Hugging Face, with download scripts (download_training.py using boto3) and tooling provided at github.com/castacks/tartanair_tools. It served as the official dataset of the CVPR 2020 Visual SLAM Challenge (monocular and stereo tracks). The dataset is released under CC-BY-4.0; the accompanying tartanair_tools software is separately BSD-3-Clause licensed.