Re-Training Basil Faulty the DonkeyCar

Basil Action Shot

I could finally set up the training environment using Ubuntu in VirtualBox on Windows to allow Basil to learn to race around the track in my kitchen!

Data Gathering Take 2

My last attempt to train Basil Faulty the DonkeyCar was only somewhat successful. He still couldn’t go around the track I trained him on so this time I decided to have a more methodical approach to the data-gathering.

I re-laid one corner of the track that was particularly tricky, even just while I was training him and deleted the old tub so none of the old tub data would taint the new neural network.

cd play/ohmc_car
rm -r tub

I also had to re-enable the joystick in


I did a few practise runs around the track without recording the data to make sure I wouldn’t be giving him the wrong information.

Training Take 2

This time when I was back on the main pc, rsync worked! Maybe this had to do with the virtual environment setup but copying the tub was so easy this time:

source ~/virtualenvs/donkeycar/bin/activate
cd ohmc_car/
rm -r tub
rsync -av pi@basilfaulty.local:play/ohmc_car/tub .

I could then train the neural network again the same way python train --tub tub/ --model models/take2.hdf5

The neural network again took approximately 28 epochs to train the model, I copied this back to the pi using scp this time:

scp take2.hdf5 pi@basilfaulty.local:play/ohmc_car/models
python drive --model ~/play/ohmc_car/models/rae_one_way.hdf5

He could then drive around the whole track (albeit slowly)! Success!

I tried increasing the speed but he was very wobbly and did go off the track a few times.

Take 3!

I did all the data-gathering and training again but this time I focussed on having a near-constant throttle and instead of ‘twitching’ the steering as I went around the track I tried having a consistent steer around a corner in an attempt to get rid of some of Basil’s wobbles.

This was very successful - after retraining, Basil could smoothly take the whole track!

About Me

Engineer, maker, do-er...
I basically just like to make things.