autonomous_car_driving

Introduction

The goal of this project was to implement the NVIDIA paper using behavioral cloning to make an End-to-End self-driving car with the Udacity simulator

Track 1

• The Udacity self-driving car simulator provides two tracks, the track 1 is the easy one. So we drive the car manually to get data(images from the cameras and steering angles)from 3 to 4 laps.

• The front cameras images will serve as input to our model and it will try to predict the respective steering angle, so it will be the output of the model

• With this data, i trained the pilotNet for 20 epochs, what gave me a good result but the car was not able to do the final right turn because of the data imbalance of the training set. In this track there are much more left turns than right turns so almost all the training examples have a negative output(steering angle), thus the probability of the model output a negative value(left turn) is much higher than a positive value(right turn). Besides, there are few turns, so the majority of the steering angles outputs are around zero and thus the model will be biased to output small values.

• To correct this data imbalance problem i used data augmentation: x and y translation, horrizontal mirroing, random shadows and bright. Check out the result by clicking in the image bellow:

Track 2

• This track is much harder compared to the track 1, because it has a lot of sharp turns and shaddows in the road. Here i modified the pilotNet replacing the elu activation by a relu and using batchNormalization in each layer. I trained this model with an one lap data and knowing that this track has several sharp turns i modified the throttle equation
  from: throttle = 1.0 - steering_angle**2 - (speed/limit)**2
  to: throttle = np.clip(1.0 - abs(steering_angle) - (speed/limit)**2, -1.0, 1.0)

so the car will desacelerate more to make a sharp turn and and i multiplied the predicted steering_angle by 1.5 to make the car turn more. Check out the result by clicking in the image bellow:

Future improviments

Somethings that might improve the model

• Try different architectures with Residual Learning or Recurrent neural networks
• Include the velocity and throttle in the network
• Get more data for the second track
• Try Reinforcement Learning