CS81 Lab1: Neural Network Brains for Robots

In this lab you will familiarize yourself with pyrobot, a tool for controlling both simulated and physical robots, that is written in python. You will also learn how to use the conx library for experimenting with neural networks.

You will go over an extended example demonstrating how you collect data for a robot task, train a network, and then test the resulting weights on the robot. You will then create a robot with a neural network brain and train it to do phototaxis (moving in response to light).

• Read through and try the commands in the pyrobot overview.
• Read through and try the commands in the conx overview.

One method for teaching a neural network to control a robot is called offline training and involves the following steps. First hand code a robot controller. Use this controller to collect data to be used to train a neural network. Next train a neural network with the collected data. Finally use the trained network to control the robot and test its performance. Follow the steps below to train a neural network to control a robot to do a wall following task.

• I have provided a hand coded program called wallFollow.py that causes the robot to follow walls on its left hand side. To observe a robot using this program do the following:

  %pyrobot -s PyrobotSimulator -w Tutorial.py -r PyrobotRobot60000 -b wallFollow.py
  

  The -s flag specifies which simulator you would like to use. Each simulator has a number of different worlds, which are specified by the -w flag. The simulator can control multiple robots simultaneously. Each robot has a name specified by the -r flag. Finally the robot is controlled by a brain, which is specified by the -b flag.

  This will open two windows: one the pyrobot control window and the other the simulator window. In the simulator window, select the View menu and choose Trail, to see the robot's path. In the pyrobot window, push the Run button to start the robot. Every 200 steps the program will report the average distance of the robot from the wall on its left side. We will use data collected from a program running a controller like this to train a neural network.

  To quit, first press the Stop button in the pyrobot window to suspend the brain. Then go to the Robot menu and select Unload Robot. Finally, go to the File menu and select Exit. If the pyrobot window fails to close, then in a terminal window do: killall -9 pyrobot.

• Open the wallFollow.py program in an editor. Notice that it contains a class that inherits from a Brain class. In pyrobot all controllers are constructed in this manner. Each brain can optionally use the setup method for initialization. It will be run once when the object is instantiated. Each brain must have a step method. This is the what will be executed on each time step (typically about 10 times a second). The move command is what causes the robot to move. It takes two parameters: translation (forward/backward) and rotation (left/right). Each of these parameters can be in the range -1 to +1.
• We will use the program collectData.py to execute the wall following behavior and collect data along the way. It is important to remember that all data used with a neural network must be scaled to match the range of the activation function. In our case, all data must be scaled between 0 and 1. To run the data collection program do:

  %pyrobot -s PyrobotSimulator -w Tutorial.py -r PyrobotRobot60000 -b collectData.py

  When you see a message that the data collection is done. Stop the robot and exit from pyrobot. You should now have two files called sensorInputs.dat and motorTargets.dat saved in your directory.
• Next we will use the program trainNetwork.py to train the neural network on the saved data. We do not need to use the simulator to do this part.

  %python trainNetwork.py
  

  This will take several minutes to complete. It will execute 1000 epochs of training on the collected data. Watch the error and percent correct during training. Does error consistently go down? Does percent correct consistently go up? After training is complete, you will have a new file in your directory called wallFollow.wts containing the weights and biases of the most successful version of the trained neural network. You will also have a log file showing the progress of learning throughout training.
• We can now use the saved weights to see how well the network controls the robot in the world it was trained on.

   
  %pyrobot -s PyrobotSimulator -w Tutorial.py -r PyrobotRobot60000 -b testNetwork.py

  Be sure to "View" the robot's "Trail" in the simulator window before pressing the "Run" button in the pyrobot window. How does the neural network controlled robot perform compared to the hand-coded teacher program? In what ways does the neural network controlled robot behave differently from the teacher?
• One claim that is made about neural networks is that they generalize well to new situations. This network was trained on data collected in the Tutorial.py world. Let's see how well it works in a new world. Try the same command as above but replace Tutorial.py with the new world LoopingMaze.py. How does the network perform in this new environment?

You will go through the same series of steps as above, but this time using your own hand-coded program for teaching a robot to orient and approach a light source.

You should use the Braitenberg.py world as the basis for your task. When collecting data be sure to start the robot from many different initial conditions (with the light straight ahead, to the right, to the left, behind the robot). You can use the Simulation device, which is discussed in the pyrobot overview, to programmatically reposition the robot after it has successfully reached the light source.

Put all of the files you created in the cs81/labs/1 directory. In the SUMMARY file provided, describe what your teacher program does, how you trained the network, and your results.

When you are done, run handin81 to turn in your completed lab work.