DeepPicar is a low-cost autonomous RC car platform using a deep convolutional neural network (CNN). DeepPicar is a small scale replication of NVIDIA's real self-driving car called Dave-2, which drove on public roads using a CNN. DeepPicar uses the same CNN architecture of NVIDIA's Dave-2 and can drive itself in real-time locally on a Raspberry Pi 3.

The release branch can be cloned using the command:

$ git clone --depth=1 https://github.com/heechul/picar/

DeepPicar is comprised of the following components:

• Raspberry Pi 3 Model B: $35
• New Bright 1:24 scale RC car: $10
• Playstation Eye camera: $7
• Pololu DRV8835 motor hat: $8
• External battery pack & misc.: $10

For manual control:

$ python picar-mini-kbd-drv8835.py

The controls are as follows:

• 'a': drive the car forward
• 'z': drive the car backward
• 's': stop the car
• 'j': turn the car left
• 'k': center the car
• 'l': turn the car right
• 't': toggle video view
• 'r': record video

For autonomous control:

$ python picar-mini-dnn-drv8835.py

Before training a model, the following changes should be made:

Change model (folder) name:

save_dir = os.path.abspath('...') #Replace ... with a name for the model

Change if normal category is to be used:

use_normal_category = True #True = equally select center/curve images, False = no equal selection

Select epochs to be used for training and validation in the params.py file:

epochs['train'] = [...] #Replace ... with integer values used to represent epochs  
epochs['val'] = [...] #Replace ... with integer values used to represent epochs

After all of the above steps are completed, The model can then be trained by running:

$ python train.py

By default, the platforms are tested over epoch 6 (out-video-6.avi), but the epochs processed can be changed by altering epoch_ids in test-model.py:

epoch_ids = [...] #Replace ... with all epochs to be processed

Also, epochs can be processed more than once (i.e. epoch_ids = [6,6] would have the platform process epoch 6 twice).

The number of frames processed can be increased/decreased as well by changing:

NFRAMES = _ #Replace _ with the total number of frames to process

For convenience, the platforms can be fully tested by running the following scripts:

Raspberry Pi 3 / Intel UP board:

$ ./test-model_timings.sh # Run all multicore and multimodel tests
$ ./benchmark_timings.sh # Run all synthetic benchmark/co-runner tests w/ perf information
$ ./benchmark_timings_noperf.sh # Same as benchmark_timings.sh but doesn't measure perf information

NVIDIA Jetson TX2:

$ ./test-model_timings_x2.sh # Run all multicore and multimodel tests while using the GPU
$ ./test-model_timings_x2_cpu.sh # Run all multicore and multimodel tests while using the CPU only
$ ./benchmark_timings_x2.sh # Run all synthetic benchmark/co-runner tests while utilizing the GPU
$ ./benchmark_timings_x2_cpu.sh # Run all synthetic benchmark/co-runner tests while only using the CPU

The bandwidth benchmark used in the co-runner tests can be found in the IsolBench suite.

Please refer to PicarMini.md for more detailed explanations/instructions.

The DeepPicar code utilizes MIT's DeepTesla (https://github.com/lexfridman/deeptesla), which provides a TensorFlow version of NVIDIA Dave-2's CNN.

NVIDIA Dave-2 (and its CNN) is described in the following paper. https://arxiv.org/pdf/1604.07316

The paper for DeepPicar can be found at https://arxiv.org/abs/1712.08644. It can be cited using the following BibTeX entry:

@article{bechtel2017picar,
	title = {DeepPicar: A Low-cost Deep Neural Network-based Autonomous Car},    
	author = {Michael Garrett Bechtel and Elise McEllhiney and Heechul Yun},
	journal= {arXiv preprint arXiv:1712.08644},
	Year = {2017}
}