Lab4

Lab 4: Behavior Cloning

In this lab, we will use a basic behavior cloning algorithm to train a model to drive the truck.

Make sure that your repo is up-to-dated

Setup Environment

1. Open a new terminal and SSH into your robot

ssh nvidia@<IP OF YOUR ROBOT>

2. Activate ros_base environment on your robot

conda activate ros_base

3. Install Pytorch

pip3 install torch

Launch Learning Node

Open three new terminals and let's call them T1, T2, and T3

1. In T1, SSH into your robot and launch SLAM

ssh nvidia@<IP OF YOUR ROBOT>
cd ~/StartUp
./start_ros.sh <IP OF YOUR ROBOT>

2. In T2, navigate to the repo on your PC

cd <REPO ON YOUR PC>/ROS_Core
conda activate ros_base
catkin_make
source devel/setup.bash
source network_ros_client.sh <IP OF YOUR ROBOT> <IP OF YOUR PC>
roslaunch racecar_interface visualization.launch enable_routing:=false

3. Start the SLAM from RQT

4. In T3, SSH into your robot and start learning node

ssh nvidia@<IP OF YOUR ROBOT>
cd <REPO ON YOUR ROBOT>/ROS_Core
conda activate ros_base
catkin_make
source devel/setup.bash
source network_ros_host.sh <IP OF YOUR ROBOT>
roslaunch racecar_learning lab4.launch

Start Training Online

1. In the RQT, call the service "learning/start_learn" from RQT to start training.

2. In the RVIZ, use the "2D Nav Goal" to set a reference path for the robot

A loop will be generated automatically as your robot's reference path, use your controller to drive the robot along the path.

3. In T3, you will see the loss be printed out.

You can drive your robot along the reference path for a few loops, then simply let the robot stop on the track and wait the loss to converge.

4. Once the loss converges, call the service "learning/eval" from RQT to pause the training and evalute the model.

Hit the down button on your controller to start the evaluation. The robot will drive along the reference path.

5. If the robot drives well, call the service "learning/save_model" from RQT to save the model and call the service "learning/save_data" to stop the training.

Your model will be saved in folder "ROS_Core/Labs/Lab4/models" on your robot, and the training data will be saved in folder "ROS_Core/Labs/Lab4/data" on your robot.

6. If you do not like your model, call the service "learning/start_learn" again from RQT to resume training again.

Traning Offline

With data collected from the previous step, we can train the model offline using provided iPython notebook. You can train this on your own PC, which should be significantly faster than the computer on the robot.

Test the Model

You can evaluate the model trained offline by using additional parameter during the launch of the learning node. In T3, relaunch the node using

roslaunch racecar_learning lab4.launch model_path:=<PATH TO YOUR MODEL>

Task: Train your own behavior cloning policy, test it on the robot and show it to your AIs