End-to-end autonomous driving in GTA V

The dataset contains 230322 cropped 640x160 RGB road frames, alongside steering, acceleration and brake inputs normalized to [0,1] and the vehicle’s current speed extracted from the interface stored in an HDF5 format. It was collected to train agents to follow the minimap route with a zoomed-in view. The vehicle used is a black Buffalo with no upgrades applied. Data was acquired between 05:00-06:00 and 17:00-22:00 in 4 different weather conditions: Clear, Clouds, Overcast and Rain. It contains 2 parts: The first 173175 frames that encompass natural driving behaviour by capturing the entire route. The following frames, which address the inherent underrepresentations of braking and extreme steering, by collecting more data only in these specific situations with higher than normal speeds, to introduce the need for stronger braking signals and simulate emergency situations. Below we present the composition of the dataset, which shows multiple driving scenarios and their proportion: * Cruising: 52814 frames (22.9%) Includes: cruising (22.5%), low acceleration straight (0.4%) * Straight coasting: 5645 frames (2.5%) * Light braking: 8775 frames (3.8%) Includes: light braking straight (0.9%), light braking with steering (2.9%) * Slight steering adjustments: 21113 frames (9.2%) Includes: slight turning with acceleration (1.9%), coasting with steering (7.2%) * Straight driving (steady acceleration): 94059 frames (40.8%) Includes: steady acceleration (38.3%), high speed driving (2.5%) * Normal & slow turning: 15923 frames (6.9%) Includes: normal turning (3.8%), slow turning (2.9%), aggressive turning (0.2%) * Moderate & emergency braking: 31993 frames (13.9%) Includes: emergency braking (7.0%), moderate braking (6.9%)