Training navigation tasks in simulators Technical exchange

[zhw970623]

Thank you very much for your work, it is excellent.

1. There are some questions I want to ask you. In the drone navigation task, during your training, each training environment is a small grid of about 1066. This space is surrounded by walls on six sides. Why do you set these 6 walls? If you train in a relatively large scene, can you successfully complete the navigation task? If you train directly in a relatively large environment without walls at the boundary (navigation task), what should you pay attention to in order to complete the training.
2. I see that you have also verified the paper published in the navigation task in flightmare. Can this part of the code be open sourced for exchange and learning?
3. Is there a tutorial for sim2real of the navigation task for exchange?

[mihirk284]

Thank you for the kind comments. I will try to answer your questions point by point:

1. We decided to choose an environment with walls quite early and decided to stick with the choice purely for not introducing too many changes before things began working. After things began working, we ended up testing in similar environments and the configuration stayed. There generally would not be a problem if you remove the walls and distribute the obstacles everywhere in the environment. One problem then becomes how many obstacles you want to simulate in total because that affects rendering speed. Also, for us it could be a bit difficult to control the difficult level (curriculum) and maintain a reasonable spread of the obstacles over the larger area when sampling.
2. Unfortunately we have upgraded operating systems and framework since the paper. I have struggled to get it working on my machine, but to the best of my knowledge, as long as you are able to interface this code (https://github.com/ntnu-arl/aerial_gym_simulator/blob/main/aerial_gym/sim2real/sample_factory_ros_node.py) to your working installation of Flightmare, it should work as expected. Please make sure the image topics, commands and states are correct but that's all that's needed to get it running.
3. There currently isn't a detailed tutorial beyond this script: (https://github.com/ntnu-arl/aerial_gym_simulator/blob/main/aerial_gym/sim2real/sample_factory_ros_node.py). I can prepare one if you point out to me the challenging aspects. Would you be interested in details for training the policy, or for transfer to a real-world platform? Overall, we assume we have MAVROS, a depth camera (and state estimation solution) installed and working on the robot for communicating state and velocity commands. If you let me know what are the confusing aspects, I can provide additional details :)