spot
spot is an open source driver assistance system for handmade toy car.
Hardware
- 4WD Smart Car Chassis
- ESP32 - system on a chip microcontroller with integrated Wi-Fi
- Mobile phone camera
Self-driving model
Self-driving model architecture is inspired by Nvidia Network from 2016.
Model is trained fully end-to-end.
It takes an image and outputs how you should drive, in our case the model outputs a steering angle value [-1, 1]
Model starts with a block of 5 fully-convolutional layers that reduces down to a 1680-dimensional vector.
This block represents our vision model and it should extract all features from an image that are essential to driving.
The resulting 1680-dim feature vector then goes into a block of 4 fully connected layers that represent our policy model.
In total model has about 240 000 parameters.
Data
Training data was collected by sampling video from mobile phone camera at 10 FPS because a higher sampling rate would include images that are highly similar, and thus not provide much additional useful information.
About 10 minutes of driving data was collected (~ 6000 frames, 90% is used for training and 10% for validation).
Results
Model has successfully learned to drive on a track he has never seen before, neither on the training dataset nor on the validation dataset.
