pjhae / Mujoco_autonomous_drone_landing

Drone landing on moving target

Autonomous Landing of Quadrotor on a Moving Target Using Vision-based RL

0. Requirements

Installation : GYM, MUJOCO, Stable-Baselines3 + (Linux)

1. Move to :

YOUR_PATH/python3.X/site-packages/gym/

2. Clone this repository in

YOUR_PATH/python3.X/site-packages/gym/envs/

3. Move [train_test_] directory to :

YOUR_PATH/python3.X/site-packages/gym/

4. Train/Test

cd YOUR_PATH/python3.X/site-packages/gym/train_test_
(Training v1) python PPO_train.py
(Training v2) python PPO_train_custum-v3.py
(Test) python PPO_check.py

1. Problem Statement

Our main Our fundamental goal is to control the robot in simulation using a vision-based RL algorithm.

TASK : Autonomous Quadrotor landing on a moving platform

Control Input(=Neural network output) and Constraints

2. Hardware 3D design (Quadrotor , Car)

Design Tool : Solidworks (+ SWtoURDF)

3. MUJOCO camera sensor

For mounting Camera on Robot Model, you can see the file in gym/mujoco/assets/Drone_ver_1.0/assets

To get RGB data from camera for observation, you can see the file in gym/mujoco/drone_v2.py

4. Policy Net

• Input : Image RGB data + Current Action and Pitch angle [Vx, Vy, Vz, Pitch angle]

• Output : Vx, Vy, Vz, Wy

• RL algorithm : PPO

5. Training env

you can see the code for specific MDP setting(S,A,R..) info

YOUR_PATH/python3.X/site-packages/gym/envs/mujoco/

drone_v1.py : Simplified vector verification

drone_v2.py : TASK

6. Turn-Off Flag

If the drone's four points touch the landing box, the propeller no longer needs to rotate

• So, Let’s set all control inputs to zero. (= Turn off)
• This algorithm is implemented using the Turn-off flag.

7. Results - Simplified vector obs verification

8. Results - Camera obs + CNN feature extraction

9. VIDEO

Video : YOUTUBE link

You can watch video through the link above.