OHMC2020 Software instructions
Contents
Overview
You will need to bring your own laptop to the assembly session.
To make the most of your time in the assembly workshop, it's best to install some tooling so you can maximise your time spent building and playing, rather than installing and configuring.
Why do we need a laptop?
From a software perspective, the DingoCar (Raspberry Pi) is self-contained ... driving, data acquisition (for training) and ultimately self-driving are all performed with on-board software. The provided micro-SD card already has all the required software pre-installed, as well as two pre-training A.I / Machine Learning models. The DingoCar software includes a web-server that provides a web interface that works on both desktop and mobile web browsers (which great for driving).
But you need a laptop to train the Neural Network ... using the data acquired on the DingoCar. To do this, you need Python and TensorFlow (a Machine Learning framework) on your laptop.
Software environment: Laptop / Desktop
For easiest, and fastest operation, you can use software on Google's Colab to process your car data in the cloud. This means all you need to have on your computer is:
- ssh: to connect to your car. (Windows users, it's now an optional Microsoft update or install putty).
- scp: to copy files to and from your car (Windows users: putty comes with pscp which is equivalent)
- a google drive account
Google Colab instructions (make this more explicit)
This is preferred because it's fast, doesn't require a big setup time on your local computer, and gives all workshop participants more time playing with their car and less time wrestling with installation and configuration issues. Also, if you don't have a GPU, it's significantly faster using Google's computer power to generate the model than it is to do it on your own.
- Go here once logged into to your google account. https://colab.research.google.com/github/robocarstore/donkey-car-training-on-google-colab/blob/master/Donkey_Car_Training_using_Google_Colab.ipynb
- Click the 'copy to drive' to make your own copy.
- Fill in the bits in the main area, step by step to follow the instructions to get the data from your car up to google, trained on google, then copy the results back down to your car. LIKE MAGIC
DIY instructions for when you don't/can't use Google Colab (move this to separate page)
We don't want Python 2, and people have reported problems with Python 3.7 or later. So we currently use Python 3.6.
Miniconda instructions:
- Go to the Miniconda archive
- Download Miniconda3-4.5.4 in the right system for you.
- In your command line prompt, go to the directory holding the file
- Run the script: ./Miniconda3-4.5.4-Linux-x86_64.sh (or equivalent)
This will by default add the Miniconda directory to your path. Now you can check you have Python 3.6 installed and available:
- python3 -i
- This should show you Python 3.6.5 | Anaconda Inc.
- Use quit() to get out of the python shell
Get Dingocar
Go to which directory you like to keep your coding projects in.
- git clone https://github.com/tall-josh/dingocar.git
- cd dingocar
- git checkout master
Install Tensorflow for machine learning
- Ubuntu
- apt-get install -y virtualenv # Note: You may be using a different software installer
- mkvirtualenv donkeycar -p python3
- pip install tensorflow==1.8.0 # Note: Probably requires Python 3.5 or 3.6. People are having problems with Python 3.7
(if you get errors, you can try (re-) installing pip: python -m pip install --upgrade pip )
- Debian, if virtualenv isn't there, try this instead
- virtualenv donkeycar -p python3
- cd donkeycar
- export PATH=`pwd`/bin:$PATH
- pip install tensorflow==1.8.0
(This seems to be v2, but 2019 instructions use 1.8?)
- conda install tensorflow-cpu
Install Dingocar
- pip install -e ./dingocar
Create an instance for your specific car
- donkey createcar --path ~/mycar #give your car its own unique name here!
Find your car
Every car has a unique hostname, from ohmc_01 to ohmc_32. The micro-SD card adapter is labeled with your car name.
Your DingoCar has been pre-configured to connect to the LCA2020 network. But how do we find out its IP address? There is a cron job that every minute sends an MQTT message to test.mosquitto.org with your car's name, IP address and a timestamp.
Install an MQTT client, as follows ...
- apt-get install mosquitto-clients
Use the following command to read the MQTT messages ...
- mosquitto_sub -h test.mosquitto.org -t 'ohmc/#' -v
ohmc/ohmc_01 10.193.2.69 Mon 11 Jun 07:08:01 UTC 2018
The IP address returned will tell you how to connect to your car. Now you can:
``ssh pi@<IP_ADDRESS>`` with the default password '`raspberry'`. Change this once you log in, using ``passwd`` command.
Driving your car manually
Caution: Put your car "on blocks" (wheels off the ground) the first time you try driving it
- ssh pi@<IP_ADDRESS>
- cd ~/mycar
- python manage.py drive
loading config file: /home/pi/play/roba_car/config.py config loaded PiCamera loaded.. .warming camera Starting Donkey Server... You can now go to http://xxx.xxx.xxx.xxx:8887 to drive your car.
With a desktop web browser, the user interface provides a virtual joystick (right-hand frame) that you can use to drive the car ... altering the steering and throttle values.
The mobile web browser, the user interface allows you to drive by tilting the phone left-right for steering and forwards-backwards for throttle. For safety, you must press the [Start Vehicle] / [Stop Vehicle] toggle button to enable control.
Further DonkeyCar docs on driving
Training your car
Once you are driving your car confidently around a track, it is time to acquire training data for the Neural Network. DingoCar operates at 10 frames per second, capturing a 160x120 image, along with steering angle and throttle value. This is all stored in the $HOME/ohmc_car/tub/ directory.
Before training, clean out previous data. Don't remove all the files in the tub/ directory: the tub/meta.json file is important.
Perform the same commands as for manual driving:
- ssh pi@$IP_ADDRESS
- cd ohmc_car
- python manage.py drive
Then via the web browser press the [Start Recording] button, drive the car around a track, then press the [Stop Recording] button.
It is recommended that you collect between 5K and 20K frames. At 10 frames per second, that is between 500 and 2,000 seconds of driving. Make sure that you drive clockwise and anti-clockwise.
You will need to type this command just once on your DingoCar to provide a directory on your laptop for your training data ...
- rsync -av pi@<car_ip>:ohmc_car .
When finished acquisition, then transfer the data from the DingoCar to your laptop / desktop for training the Neural Network.
- rsync -av pi@<car_ip>:ohmc_car/tub ohmc_car/tub_$DATE
Once training data has been copied to your laptop / desktop, you can begin training the Neural Network.
Extensive DonkeyCar documentation
Run these commands on your laptop / desktop to train the Neural Network ...
- workon donkeycar # For those who have set-up a virtualenv
- cd ohmc_car
- python manage.py train --tub $HOME/ohmc_car/tub_$DATE --model ./models/model_$DATE.hdf5
using donkey version: 2.5.7 ... loading config file: /Users/andyg/play/ai/roba_car/config.py config loaded tub_names ./tub_2019-01-15c train: 5740, validation: 1436 steps_per_epoch 44 Epoch 1/100 2019-01-21 13:08:49.507048: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA 43/44 [============================>.] - ETA: 0s - loss: 58.5130 - angle_out_loss: 30.3421 - throttle_out_loss: 86.6839 Epoch 00001: val_loss improved from inf to 0.19699, saving model to ./models/roba0_2019-01-16c.hdf5 44/44 [==============================] - 38s 874ms/step - loss: 57.1887 - angle_out_loss: 29.6601 - throttle_out_loss: 84.7172 - val_loss: 0.1970 - val_angle_out_loss: 0.3230 - val_throttle_out_loss: 0.0710
On a modern laptop, each epoch will take around 30 seconds to complete. For up-to 100 epochs. Typically, you can expect around 20 to 40 epochs before the Neural Network stop learning. That is around 10 to 20 minutes of training time.
The training command creates the Neural Network weights that represent what your DingoCar has "learned".
Once it's done, copy the trained model back to your DingoCar (Raspberry Pi).
- scp $USERNAME@$HOSTNAME:ohmc_car/models/model_$DATE.hdf5 models
Further DonkeyCar info on training
Letting your car drive itself!
Once your trained model has been copied back onto the DingoCar, your car can be self-driven as follows:
- python manage.py drive --model ~/ohmc_car/models/models/model_$DATE.hdf5
This works similar to the manual driving mode with the addition of a trained model that can either:
1) User: Manual control of both steering and throttle 2) Local Angle: Automatically control the steering angle 3) Local Pilot: Automatically control both the steering angle and throttle amount
The web browser provides a drop-down menu to select between these options.
It is recommended to just start with "Local Angle" and control the throttle manually with the "i" key (faster) and "k" key (slower).
More information, for later
There's extensive DonkeyCar documentation if you're looking for more detailed installation or configuration instructions.
Background information
Software installation: Raspberry Pi
Your DonkeyCar (Raspberry Pi) is already pre-installed. This section is for reference only.
Extensive DonkeyCar documentation
- sysctl -w net.ipv6.conf.all.disable_ipv6=1 # Reconnect via ssh afterwards
- sysctl -w net.ipv6.conf.default.disable_ipv6=1
- apt-get update
- apt-get upgrade
- apt-get install -y vim git mosquitto-clients
- apt-get install -y virtualenv build-essential python3-dev gfortran libhdf5-dev libatlas-base-dev libopenjp2-7-dev libtiff5
- apt-get install -y i2c-tools
- i2cdetect -y 1
0 1 2 3 4 5 6 7 8 9 a b c d e f 00: -- -- -- -- -- -- -- -- -- -- -- -- -- 10: -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 20: -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 30: -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 40: 40 -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 50: -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 60: -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 70: 70 -- -- -- -- -- -- --
- mv env env.donkeycar
- virtualenv env -p python3
- pip install tensorflow==1.8.0
- pip install adafruit-pca9685
- pip install picamera
rPi reboots and Auto-starting donkeycar without logging in
If your rPi reboots when you're driving, check the following things:
- did you swap the servo wires between steering and power. Turn the ESC off, unplug one servo wire, turn the ESC back on (little toggle switch), and test steering. Make sure that steering does not activate the wheel motors.
- is your battery low? are you accelerating too quickly?
Because the battery has a high internal resistance, if you demand a lot of power too quickly, the voltage on the battery drops enough that the rPi gets below its critical voltage (when you add the voltage drop from the voltage converter), and the rPi reboots.
To restart donkey server automatically, do the following on your car:
- Add this to /etc/rc.local:
su - pi -c 'cd ~/ohmc_car; python manage.py drive &>/tmp/out' &
- move 'source ~/env/bin/activate' from ~/.bashrc to ~/.profile
Steering and throttle calibration
To save time at the workshop, you won't need to calibrate your car's steering and/or throttle. However, you may get better results and can perform calibration when you have time.
DonkeyCar instructions on calibration
OLED digital display
Something here about image/text display for the car info.