Unity-based driving simulator

This repository contains a Unity-based driving simulator for training reinforcement learning agents.

Features

• Domain randomization (textures, cubemaps, driving dynamics, track properties, random objects around the track)
• OpenAI Gym compatible

Getting started

Build the project with Unity Editor. Make sure that DrivingLearningBrain is selected instead of PlayerBrain from the Academy component as well as the Agent component. To use the compiled binary from Python, install gym-unity and ml-agents packages with Pip:

$ pip install gym-unity ml-agents

Check out gym_example.py for an example on how to run the simulator:

from gym_unity.envs import UnityEnv

env = UnityEnv('build', worker_id=0, use_visual=True)

for _ in range(10):
    env.reset()
    done = False
    while not done:
        action = env.action_space.sample()
        obs, reward, done, _ = env.step(action)
        print(reward)

If you want to pass command line arguments to the simulator when using Python, you can make a wrapper bash script that is passed to the UnityEnv, which contains the arguments needed:

#!/bin/bash
./carsim.x86_64 --segmentation $@

Save the wrapper as e.g. 'wrapper.x86_64' (UnityEnv requires the file extension)

Configuration

There are various configuration options that can be altered in the Editor. The Academy component contains the following options:

• Width : Observation width in pixels
• Height: Observation height in pixels
• Quality Level: Rendering quality
• Time Scale: Speed at which the simulation is run. There are some problems with higher values (over 10)
• Target Frame Rate
• Reset Parameters: These parameters control the probabilities ([0, 1]) of randomizing specific aspects on episode reset. The default values can be set here, and they can be altered at run-time from the Python interface.

In addition, the Agent component contains the following options:

• Decision Interval: Action repeat, i.e. how many steps should one action be performed for before requesting a new one

Reset parameters

Note: this feature requires a patched version of gym-unity, which can be found here.

Reset parameters allow run-time modification of randomization probabilities for different aspects of the simulator. Each parameter is a value between 0 and 1, and represents the probability of randomizing that aspect when the environment is reset at the end of an episode. The environments reset function takes a dict of Reset Parameters as a keyword argument. For example:

env.reset(reset_params={'random_track': 1., 'random_texture': 1., 'random_fov': 1., 'random_cubemap': 1.})

List of available parameters:

• random_track: Randomize track shape and width
• random_texture: Randomize all textures (track, ground, etc.)
• random_fov: Randomize camera field of view
• random_cubemap: Randomize skybox cubemap
• random_camera_height: Randomize camera height from ground
• random_dynamics: Randomize various car dynamics (torque, steering angle, top speed, etc.)

Shaders

The simulator contains a couple useful shaders: a semantic segmentation shader, and a depth shader. These can be enabled with --segmentation and --depth command line arguments respectively.

Recording data

Collecting observation frames from the simulator can be enabled with --save-frames command line argument. In addition, these arguments can be used to configure the data collection:

• --save-every n: save every n:th frame
• --save-width x: image width for the saved frames
• --save-height x: image height for the saved frames

The data is saved in the current working directory, to a new directory called dataset.