Domain Randomization for Transferring Deep Neural Networks from Gazebo to Real World Using ROS
The Construct Robotics Institute 4:14
1,658 views · 13 likes Watch on YouTube ↗
We have replicated the results of the amazing paper by OpenAI "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World". We use ROS and Gazebo for the whole simulation, control and cameras, as well as a special Gazebo plugin we have created to implement the domain randomization.
The whole code is going to be provided for free as a ROSject (http://www.rosjects.com), which means that you will have access to the full source code of the whole experiment. On top of that: you will be able to reproduce the same results from minute 1, because the ROSject ensures that it will work for you in the same way it works for us. So you will be able to reproduce the results and start working towards improving it without having to start building again from scratch.
The ROSject contains:
* The simulated training and testing environment
* The domain randomization plugin for Gazebo which changes lights, colors and locations of objects randomly
* The training method for the DCNN: how to capture training data, and how to use it to train the model.
* The code that implements the whole pipe of testing the trained DCNN for the control of the Fetch robot
* A document explaining how everything works and how to proceed to launch training and execution.
* Everything is ROS based, so reproduction on the real robot is ensured
* Everything reproducible in ANY computer
We are still finishing some details, but we are going to provide the ROSject in a few days. If you want to get the ROSject, please like this video and leave us a message in the video comments area below. We will contact you in a few days with the ROSject.
* The article we are reproducing: https://arxiv.org/abs/1703.06907
* The ROS Development Studio: http://rosds.online
* ROSjects for robotics code reproducibility: http://www.rosjects.com
The whole code is going to be provided for free as a ROSject (http://www.rosjects.com), which means that you will have access to the full source code of the whole experiment. On top of that: you will be able to reproduce the same results from minute 1, because the ROSject ensures that it will work for you in the same way it works for us. So you will be able to reproduce the results and start working towards improving it without having to start building again from scratch.
The ROSject contains:
* The simulated training and testing environment
* The domain randomization plugin for Gazebo which changes lights, colors and locations of objects randomly
* The training method for the DCNN: how to capture training data, and how to use it to train the model.
* The code that implements the whole pipe of testing the trained DCNN for the control of the Fetch robot
* A document explaining how everything works and how to proceed to launch training and execution.
* Everything is ROS based, so reproduction on the real robot is ensured
* Everything reproducible in ANY computer
We are still finishing some details, but we are going to provide the ROSject in a few days. If you want to get the ROSject, please like this video and leave us a message in the video comments area below. We will contact you in a few days with the ROSject.
* The article we are reproducing: https://arxiv.org/abs/1703.06907
* The ROS Development Studio: http://rosds.online
* ROSjects for robotics code reproducibility: http://www.rosjects.com
Playback is via YouTube's official embedded player. Data from YouTube; Exumo is not affiliated with YouTube.