Autonomous_Driving_in_CARLA

Introduction:

We all know that world of automated systems are going to rule this world in coming times. So, here is our project on autonomous driving using Carla.

Dataset Used

We used the official dataset provide by carla team, which is of 24 GB by size. Dataset - link

Imitation Learning:

We used the method of Imitation learning to make the car an automated system in the given scenarios.

Scene	Sample Image
Cruise Control
Road Disturbance
Dual Disturbance
Multiple Disturbance

Models architecture

We used four different models for this purpose, the models which one want to use can be control by the argument "--scenario" as shown below

For model architcture, you can go this link

Loss Function

We have defined loss function for scenario==1 as the Mean Squared Error loss for each branches, speed controller and uncertainity controller. While for other cases we did following as shown in image.

Running the Code:

To run the code, bash this command on the terminal. You can change scenario as per your requirement.

            $ python main.py    
            --batch-size 1000
            --scenario cruise_control
            --workers 24
            --speed-weight 1
            --learning-rate 0.0001
            --lr-gamma 0.5
            --lr-step 10
            --train-dir "/path/to/the/training_data"
            --eval-dir "/path/to/the/eval_data"
            --gpu 0
            --id name_of_policy

Check the training log through tensorboard.

          $ tensorboard --logdir runs

References

Visual-based Autonomous Driving Deployment from a Stochastic and Uncertainty-aware Perspective