Update: I'm having to work on a different project for my master's so this will be indefinitely suspended. I hope to return to it in future.
My plan is to train a Jumping Sumo minidrone from Parrot to navigate a track using reinforcement learning. This project will be divided into several stages:
- Implement the ARSDK3 protocol in python to allow me control the drone directly via a PC and stream video as well
- Implement Deep Deterministic Policy Gradient (DDPG) in CNTK (maybe Tensorflow?)
- Use DDPG to train the drone to navigate a paper track
- The higher the speed of the robot, the greater the reward
- Negative rewards for leaving the track. I plan to train a convolutional network to decide whether or not the robot is within the track
- At each time step, the state will be the last frame (or stack of frames) as well as last read speed from the robot
- The agents actions will be speed [-100, 100] and turn [-100, 100]
More requirements will be added as the project progresses.
- Python 3
- OpenCV
- One Jumping Sumo (I am using a Jumping Race Max, other Jumping Drones should also work)
- Several batteries 😩
- I was able to install OpenCV 3 in my conda environment using:
conda install -c menpo opencv3
- The minidrone module was adapted from forthtemple/py-faster-rcnn and haraisao/JumpingSumo-Python . As I am only interested in ground motion, I have limited my implementation to that (no jumping, ...)
- The Parrot ARSDK3 document can be found here. I was also able to find some useful information on their GitHub page.