Handling dynamic obstacles

[20170505]

Hello author, thank you for your excellent open-source work.
While reading about your method, my understanding is that the local planner is trained in a static environment.
Therefore, I would like to ask whether it would be reasonable or feasible to introduce dynamic obstacles (pedestrians or randomly moving geometric shapes) into the training environment to enhance the local planner ability to handle dynamic obstacles?

[reiniscimurs]

Hi. Yes, you can but that alone will not produce the desired effect. There is no internal state of the transitions or history given to the model at any given step. So what the model sees is simply a snapshot of the environment, even if the environment is dynamic. While the model might learn to give more deference to obstacles, it has no way to avoid them in an intelligent way. For that, you either need to give history information to the model (multiple timesteps as a sequence) or use some internal state through RNN implementations.

For RNN methods, you can see the implementation in (based on this repo): https://github.com/Barry2333/DRL_Navigation
Or derived implementation in (in a different simulator, but fulfilling the same task): https://github.com/reiniscimurs/DRL-robot-navigation-IR-SIM

However, even in deployment, even the current method works quite well due to its quick reaction time even for a single state information as can be seen in: https://youtu.be/MhuhsSdzZFk?si=PPjEELg0jFgLemqG&t=260