aslteam
In this video, we demonstrate a method to control a quadrotor with a neural network trained using reinforcement learning techniques. With reinforcement learning, a common network can be trained to directly map state to actuator command making any predefined control structure obsolete for training.
More detail regarding the paper can be found from https://arxiv.org/abs/1707.05110
and implementation is available from https://bitbucket.org/leggedrobotics/rai
Source
Good job guys. Can you share details of the platform you're using – is the machine learn platform tensorflow/python-based, and is the quadrotor system running on ROS?
This looks amazing!
I have experience with ROS and tensorflow (very basic). I'm highly interested in applying your findings to a quadcopter I'm working on.
Can you please point me in the right direction?
What can I start with and what must I learn?
Thanks!
You guys are amazing, thank you for showing the direction
What simulation engine did you used? Was that Unity? How did you dealed with reality-gap – what was your approach here. Did the copter flight on policy trained only in simulation?
Sir!
Which simulator have u used for training?
Cool work! I just completed a similar project, but I used DDPG with SNNs, nice to see implementation on real quadcopter, have you tried different tasks?
mY GOSH!
I've read the paper however I could not understand how do you define the 4 actions? Each action is each rotor's velocity, right? But then how do they choose the velocity in the beginning?
Random question here: I'm doing something similar using PPO on a quadrotor (simple simulation using OpenAI gym). I'm trying to get the cable suspended load case now, but still struggling with the reward function. Terminal states are way more difficult to handle than the standard quad.
The performance of the quadrotor is excellent.
But i have a question about the reinforcement learning. I want to know what a role the RL algorithm play in the control system? navigation or attitude control?
RL algorithm just provide where to move, and PID controler to finish attitude control? Or both attitude control and location control are all RL algorithm`s duty?
Thanks for your response.
These 4 PWM actions can be perform at the same time in one iteration loop?
Hello, What a great video. I'm new to the subject of Reinforced Learning and hoped I could ask a question. Here goes: Does the ability of the drone to recover depend on input from the various sensors e.g. gyro. So say the drone flew from shade to sunlight, would there need to be a temperature sensor on board to allow it to cope with any sudden movement associated with moving from hot to cold. Or does the reinforced learning model not care about any of that and just learn to deal with whatever it encounters (so you could throw away the temperature sensor) and it would stabilize just fine with any sudden movement associated with hot to cold. Thanks
what drone did you use? i planning to buy a drone for my research to do similar things. can some suggest me a drone
Thank you very much, that is really awesome to see
Hello good job, excuse me I'm a beginner and I don't understand why the value network is essential for learning. With only one value of distance between the quad and the point this could not be enough.
This is great indeed
where is fucking source code
I have a lot of simulation files for four rotors，If someone needs can contact me by email：wujingwei1995@gmail.com
Amazing work! How did you guys handle the noise in sensors (Gyro/Accelerometer) ?