Using DDPG and ConvLSTM to control a drone to avoid obstacle in AirSim

This project is developed based on Microsoft-Airsim, achieving obstacle avoidance in height direction.

python == 3.6.2

AirSim == 1.2.0

Keras == 2.1.6

msgpack-rpc-python 0.4.1

numpy == 1.16.0

tensorflow == 1.8.0

pillow == 8.4.0

opencv-python == 3.2.0.7

• batch size is 32
• Input of the image branch is previous 4 moments and current 64x64 depth images
• Input of the state branch is previous 4 moments and current height states
• Output of the actor network is current height control instruction

• batch size is 32
• Input of the image branch is previous 4 moments and current 64x64 depth images
• Input of the state branch is previous 4 moments and current height states
• Input of the action branch is previous 4 moments and current actions
• Output of the critic network is Q value

1. run AirSim to start the simulation enviroment
2. run main.py to train it
3. run test.py to test it

I just did it for a course design task of Artificial Neural Network and Control, a master degree course of Harbin Institute of Technology, Shenzhen. So there are certainly many bugs and I didn't finetune it. If you want to do some research based on it, you could contact me and we could do it together!