Robot Obstacle Avoidance Controller Based on Deep Reinforcement Learning

Algorithm 1

DQN.

Input: Pixels and reward

Output: Q action-value function

Initialization

Initialize replay memory space

Initialize the Q network (action-value function) with random weights

Initialize target network (action-value function) with weights

1: Fordo

2:  Initialize sequence and preprocessed sequence

3:  Fordo

4:   Following policy, select

5:   Run action in an emulator and observe the reward and image

6:   Set and preprocess

7:   Store transition in

8:   Sample random minibatch of transitions from

9:   Set

10:   Calculate the loss (Perform a gradient descent step on)

11:   Train and update weights of

12:  End

13: End