Review of current neural architectures used

less than 1 minute read

These are the current 4 neural approaches explored for controlling the robot.

Pilotnet

Pilotnet architecture
Pilotnet rachitecture
  • BatchNorm
  • Conv2D x 5
  • Flatten
  • Dense x 5

Deepest LSTM TinyPilotnet

Deepest LSTM TinyPilotnet architecture
Deepest LSTM TinyPilotnet
  • Conv2D * 3
  • Dropout
  • Reshape
  • ConvLSTM * 3
  • Flatten
  • Dense * 3

Pilotnet 3D (with memory)

Pilotnet3D architecture
Pilotnet rachitecture 3D (with memory)
  • BatchNorm
  • Conv3D x 5
  • Flatten
  • Dense x 5

Deepest LSTM TinyPilotnet 3D (with memory)

Deepest LSTM TinyPilotnet 3D architecture
Deepest LSTM TinyPilotnet 3D (with memory)
  • TimeDistributed(Conv2D) * 3
  • Dropout
  • Reshape
  • ConvLSTM * 3
  • Flatten
  • Dense * 3