Demo of Autonomous driving of Prius in ROS/GAZEBO

This is a simulation of a Prius in gazebo 8 with sensor data being published using ROS kinetic The car's throttle, brake, steering, and gear shifting are controlled by publishing a ROS message. A ROS node allows driving with a gamepad or joystick.

Requirements

This demo has been tested on Ubuntu Xenial (16.04)

• An X server
• Docker
• nvidia-docker

Running directly (inference)

To have the car drive autonomously using the pre-trained weights,
Make sure the cruize_control flag is set to true here. This allows the car to drive at a constant velocity.
RUN: roslaunch car_demo demo_drive.launch This will load the Sonoma Raceway track and along with prius. Here the car will drive around the track by subscribing to /steering_command from the model prediction.

Steps for training and testing new models:

Using only the front facing Center camera for training:

Recording Bag Files and data generation

• Make sure the cruize_control flag is set to false here.
• RUN roslaunch car_demo demo_record_center.launch
• Open a terminal and navigate to car/predict_steering/dataset, Here we store all the data required for training, (i.e images and the csv files) and RUN:

rosbag record /filtered/steering_angle /filtered/image_raw 

This will save the rosbag file with images from the center front facing camera on the car along with the corresponding steering angle.

• Start driving the car using WASD or a joystick. End the recording of rosbag file when you have traversed enough in the track. Record more rosbag files (traversing the same track 4-5 times), this will allow enough data for training.
• Once we have the bag files, we have to extract them into images and steering angles. RUN rosrun dataprocess bag_extract_data_center.py, this will generate a csv file with path to the images from center camera along with the corresponding steering angles.

Training the Model

• To train the model, RUN rosrun model train_center.py , this will start training the model using the data generated from the previous process.
• After each epoch, only the best weights are saved automatically, i.e weights are saved only if there is an improvement in training accuracy from the previous epochs.

Inference (testing on the car)

• RUN roslaunch car_demo demo_drive.launch to test the model on the car.

Using the three front facing cameras of the car: Center, left and right for training:

Recording Bag Files and data generation

• RUN roslaunch car_demo demo_record_lcr.launch
• Open a terminal and navigate to car/predict_steering/dataset, Here we store all the data required for training, (i.e images and the csv files) and RUN:

rosbag record /filtered/steering_angle /filtered/center/image_raw /filtered/left/image_raw /filtered/right/image_raw 

This will save the rosbag file with images from three front facing cameras on the car along with the corresponding steering angle.

• Start driving the car using WASD or a joystick. End the recording of rosbag file when you have traversed enough in the track. Record more rosbag files (traversing the same track 4-5 times), this will allow enough data for training.
• Once we have the bag files, we have to extract them into images and steering angles. RUN rosrun dataprocess bag_extract_data_lcr.py, this will generate a csv file with path to the images from all three cameras along with the corresponding steering angles.

Training the Model

• To train the model, RUN rosrun model train_lcr.py , this will start training the model using the data generated from the previous process.
• After each epoch, only the best weights are saved automatically, i.e weights are saved only if there is an improvement in training accuracy from the previous epochs.

Inference (testing on the car)

• RUN roslaunch car_demo demo_drive.launch to test the model on the car.