Use our pretrained example agent, training code, and dataset to get started building your own AI and try to beat our high score!

Download the 100GB, 8.2 hour, dataset we used to train the baseline agent, complete with depth, DAgger labels, and capture augmentation used to achieve scores on-par with the path-follower.

Our AI, Unreal plugin, simulation assets and code, and Unreal Engine are all on GitHub, so you can modify the agent and the environment to automate labeling and iterate on ideas in minutes instead of weeks or months.

Enable up to eight cameras with images and depth transferred at blazing speeds via shared memory. i.e. Eight 512x512 48 bit RGB with 16 bit depth channel cameras, uncompressed on a GTX 980 over PCIe 3.0 x 16 are available at 20Hz as NumPy arrays.

First class support for Linux and Windows, one command python install.py, and you'll be piloting Deepdrive in minutes.

Quickly change anything in the simulation with Unreal Blueprints, using the full power of Unreal's C++ API.

Everything is C++ under the hood, so you can integrate existing systems and highly optimize anything you want.

Interface through the Gym API using a reward function based on speed, safety, legality, and comfort.

import deepdrive

def main():
    env = deepdrive.start()
    forward = deepdrive.action(steering=0, throttle=1, brake=0)
    done = False
    while not done:
        observation, reward, done, info = env.step(forward)

if __name__ == '__main__':
    main()