Raspberry Pi Remote-controlled/autonomous cars. Different approaches for control and driving are available:

• Fully autonomous through deep learning neural networks
• Driving aid, pseudo-automated (tries to avoid obstacles)
• Manual controlled via HTML interface

• robot: RC car with one motor for power, one motor for steering.

• controller: Raspberry Pi 3 B or 3 B+. Raspberry PI zero w can be used only with SKLearn and no camera.

• ADC controller: 4 channel 16bit, ADS1115 (Op. Voltage: 3.3V)

• Accelerometer: XYZ, ADXL345 (Op. Voltage: 3.3V)

• Stepper motor controller: Adafruit TB6612, 1.2A DC (Op. Voltage: 3.3V or 5V)

• Sensing the motor status requires a 1.5MOhm resistor to from the power motor leads to the ADS1115.

• Sonars: 4 HC-SR04 (left, right, center, back); (Op. Voltage: 5V) Warning: the output operating voltage is 5V, while RPi accepts only up to 3.3V, so you need a voltage divider.

• Radar (support is upcoming): 1 microwave Doppler Radar HB100

• Raspberry Pi Camera Module

• Optional: Google Coral EdgeTPU for on device efficient inference is supported. Please refer to the product page for installation.

• Deprecated IR sensors:

  • 2 front left side break beam sensors 3mm LEDs
  • 1 front GP2Y0A21YK0F Sharp IR Analog Distance Sensor 10-80cm

Since it uses Python3, it requires the following packages

sudo apt-get install python3-rpi.gpio python3-dev python3-pip libopenjp2-7 libatlas-base-dev libtiff5 python3-picamera libhdf5-serial-dev libhdf5-dev

You would then need to install two libraries for the ADC (ADS1115) and accelerometer (ADXL345)

sudo pip3 install adafruit-adxl345 adafruit-ads1x15 Pillow configparser wheel

Machine learing is carried out using Sklearn by default, which needs the following packages:

sudo pip3 install numpy sklearn 

Alternatively and only when used with a Raspberry Pi 3, TensorFlow can be used instead. To install:

sudo pip3 install tensorflow tflite-runtime

When inside teh folder Pi-bot/piRC from the command line:

python3 setup.bdist_wheel
sudo pip3 install --upgrade dist/pirc-xxxxxxxx-py3-none-any.whl  

You can also run the software directly from the folder (although it is discouraged).

• Upon booting Raspbian, the motors may not be responsive due to some GPIO channels not being properly initialized. Use (piRC/other/pirc_init.py) to initialize them.

All required parameters for ML and for the RC are included into pirc.ini which is included in Pi-bot/pirc. If not present, a default file is created upon launching the software.

This software (piRC/machine_learning/piRC_ML.py) is designed to accomplish three main tasks

• Collection mode (CM) Collect sensor data and save it into what will be a training file. When the software is operated with CM, the user is supposed to operate the RC in fully manual mode (i.e. using the original remote control. You can stop the collection by exiting with CTRL+C. The result is a file (Training_splrcbxyz.txt) with sensor readings that are ready for machine training. Examples are provided in piRC/piRC_ML/Training (it is strongly recommended to create your own). CM mode runs using this command:

  pirc -c

• Training mode (TM) Using the training file (for example Training_splrcbxyz.txt), a neural network training model is created and saved with the extension .nnModelC.pkl. Training can be conducted in the Rpi within the RC car itself, or in any other computer running python and related libraries. This usually is much much faster. TM mode runs using this command:

  pirc -t <training file>

• Running mode (RM) Once the training is completed, you can run run the inference mode (which uses sensor data acquired in real time to infer steering and power) and ultimately the RC car itself through this command

  pirc -r Training_splrcbxyz.txt

Both Classification (default) and Regression can be used.

A few notes on RM:

• If training model is not available, one is created at launch time.
• a bash script is provided (piRC/machine_learning/syncTraining/syncTFile.sh) to allow the transfer of the training file to a separate computer, run the training in that computer and retrieve the model for usage in the piRC.
• The training file can be improved while in RM by adding real time sensor data to it (the flag saveNewTrainingData needs to be set to True). This is useful as a new training model can be created on the fly (using the bash script as per the previous point). piRC_ML.py can be set to periorically reload (default is 15 seconds, change using syncTimeLimit) the training model without relaunching the script (the flag syncTrainModel needs to be set to True).
• Two Machine learning methodsare available: using the MultiPerceptron Classifier (MLPClassifier, recommended) or Regressor (MLPRegressor, highly experimental and not recommended).
• Two operation modes are available, and can be enabled using the flab runFullAuto:
  • Partial Auto: piRC will only respond to the sensor data, but it will no proactively "drive forward" if it has no reason for it. You need to use manual commands to help it.
  • Full Auto: piRC will proactively drive forward if it has been standing still for a given amount of time.

piRC can be controlled using a web interface. Communication between PHP and python is done:

• via shared file (piRC_file). Auto and manual control take place within the same python script
• via shell (piRC_nofile): Auto control is automated, in the background. Manual control is done via python script when the user is calling for an action.

• pirc_init: Initialize sensors and GPIO ports for full use with remote control.

To be working properly, a library file piRC_lib.py needs to be copied inside each of the folders:

• piRC_file
• piRC_nofile
• piRC_ML
• Upon booting Raspbian, the motors may not be responsive due to some GPIO channels not being properly initialized. Use (piRC/other/pirc_init.py) to initialize them.