SelfDrive_AI

Reinforcement learning for self-driving in a 3D simulation (Created using UNITY-3D)

1. Requirements for the SelfDrive_AI Gym

You need Python 3.6 or later to run the simulation. (Note: the current environment is only supported in windows) Also, you can directly interact with the simulation by clicking the exe file and then by using W,A, S and D keys.

Please follow the two links below to install Unity-Gym and Stable-Baselines. Also, you can train it using your custom reinforcement learning algorithms by following the OpenAI gym structure (https://gym.openai.com/).

Install Unity-Gym

Install Stable-Baselines3

mlagents can be installed using pip:

$ python3 -m pip install mlagents

The image below illustrates the target goal of the AIcar, where the car needs to explore all the trajectories to find the bridge first.

2. (Training) You can train the environment by using the code below which has OpenAI gym structure. It will save the training results into a log directory which you can view using tensorboard. Feel free to change the parameters inside the code

from stable_baselines3 import PPO, SAC, ppo
from mlagents_envs.side_channel.engine_configuration_channel import EngineConfigurationChannel
channel = EngineConfigurationChannel()
from gym_unity.envs import UnityToGymWrapper
from mlagents_envs.environment import UnityEnvironment
import time,os
from stable_baselines3.common.vec_env import DummyVecEnv
from stable_baselines3.common.monitor import Monitor
from stable_baselines3.common.policies import ActorCriticPolicy
import math


env_name = "./UnityEnv"
speed = 15


env = UnityEnvironment(env_name,seed=1, side_channels=[channel])
channel.set_configuration_parameters(time_scale =speed)
env= UnityToGymWrapper(env, uint8_visual=False) # OpenAI gym interface created using UNITY

time_int = int(time.time())

# Diretories for storing results 
log_dir = "stable_results/Euler_env_3{}/".format(time_int)
log_dirTF = "stable_results/tensorflow_log_Euler3{}/".format(time_int) 
os.makedirs(log_dir, exist_ok=True)

env = Monitor(env, log_dir, allow_early_resets=True)
env = DummyVecEnv([lambda: env])  # The algorithms require a vectorized environment to run


model = PPO(ActorCriticPolicy, env, verbose=1, tensorboard_log=log_dirTF, device='cuda')
model.learn(int(200000)) # you can change the step size
time_int2 = int(time.time()) 
print('TIME TAKEN for training',time_int-time_int2)
# # save the model
model.save("Env_model")


# # # # # LOAD FOR TESTING
# del model
model = PPO.load("Env_model")

obs = env.reset()

# Test the agent for 1000 steps after training

for i in range(400):
    action, states = model.predict(obs)
    obs, rewards, done, info = env.step(action)
    env.render()

To monitor the training progress using tensorboard you type the following command from the terminal

$ tensorboard --logdir "HERE PUT THE PATH TO THE DIRECTORY"

3. (Testing) The following code can be used to test the trained Humanoid Agent

from stable_baselines3 import PPO, SAC, ppo
from mlagents_envs.side_channel.engine_configuration_channel import EngineConfigurationChannel
channel = EngineConfigurationChannel()
from gym_unity.envs import UnityToGymWrapper
from mlagents_envs.environment import UnityEnvironment
import time,os
from stable_baselines3.common.vec_env import DummyVecEnv
from stable_baselines3.common.monitor import Monitor
from stable_baselines3.common.policies import ActorCriticPolicy
import math


env_name = "./UnityEnv"
speed = 1


env = UnityEnvironment(env_name,seed=1, side_channels=[channel])
channel.set_configuration_parameters(time_scale =speed)
env= UnityToGymWrapper(env, uint8_visual=False) # OpenAI gym interface created using UNITY

time_int = int(time.time())

# Diretories for storing results
log_dir = "stable_results/Euler_env_3{}/".format(time_int)
log_dirTF = "stable_results/tensorflow_log_Euler3{}/".format(time_int)
os.makedirs(log_dir, exist_ok=True)

env = Monitor(env, log_dir, allow_early_resets=True)
env = DummyVecEnv([lambda: env])  # The algorithms require a vectorized environment to run


model = PPO.load("Env_model")

obs = env.reset()

# Test the agent for 1000 steps after training

for i in range(1000):
    action, states = model.predict(obs)
    obs, rewards, done, info = env.step(action)
    env.render()

***Note: I am still developing the project by inducing more challenging constraints.

Reinforcement learning for self-driving in a 3D simulation

Related tags

Overview

SelfDrive_AI

1. Requirements for the SelfDrive_AI Gym

2. (Training) You can train the environment by using the code below which has OpenAI gym structure. It will save the training results into a log directory which you can view using tensorboard. Feel free to change the parameters inside the code

3. (Testing) The following code can be used to test the trained Humanoid Agent

Owner

Surajit Saikia

2021:"Bridging Global Context Interactions for High-Fidelity Image Completion"

Jiminy Cricket Environment (NeurIPS 2021)

This code implements constituency parse tree aggregation

Tensors and Dynamic neural networks in Python with strong GPU acceleration

Learning hidden low dimensional dyanmics using a Generalized Onsager Principle and neural networks

Jarvis Project is a basic virtual assistant that uses TensorFlow for learning.

Code for Multiple Instance Active Learning for Object Detection, CVPR 2021

Scientific Computation Methods in C and Python (Open for Hacktoberfest 2021)

Robotics with GPU computing

SimDeblur is a simple framework for image and video deblurring, implemented by PyTorch

Code Repository for The Kaggle Book, Published by Packt Publishing

Meta Representation Transformation for Low-resource Cross-lingual Learning

Road Crack Detection Using Deep Learning Methods

This repo is customed for VisDrone.

Code for the paper Relation Prediction as an Auxiliary Training Objective for Improving Multi-Relational Graph Representations (AKBC 2021).

Easy-to-use library to boost AI inference leveraging state-of-the-art optimization techniques.

A unified framework for machine learning with time series

[WACV21] Code for our paper: Samuel, Atzmon and Chechik, "From Generalized zero-shot learning to long-tail with class descriptors"

Codebase for the solution that won first place and was awarded the most human-like agent in the 2021 NeurIPS Competition MineRL BASALT Challenge.

ICLR 2021, Fair Mixup: Fairness via Interpolation