Conversation
… example files. Now just need to use the map...
… on what it does with obstacles
| This program will load in the task logs in `log_dir` and use them to generate actions in the `navigation2` task. | ||
|
|
||
| Information about the arguments can be found by running `python3 NavigationExpert.py -h` | ||
|
|
There was a problem hiding this comment.
Please add an example command
There was a problem hiding this comment.
This file doesn't appear to actually run the expert algorithm, rather it constructions the agent action from the expert. Do you plan to merge the expert into this file? Maybe we call this file something like navigation_select_actions_from_expert.py
There was a problem hiding this comment.
How do you anticipate we will integrate these expert demonstrations into the RLPD architecture. I imaging that the files in log_dir will act like our expert dataset, which we will then call to rollout an episode on the environment. Before pushing this to main, I think we should figure out this integration. For example, I think we could create Collector class that does "symmetric sampling", wherein it pulls have the batch from the expert and half from the policy. See the following link: https://pytorch.org/rl/reference/collectors.html
There was a problem hiding this comment.
I have verified this file runs! Good job
expert/NavigationExpert.py
Outdated
| obs = [] | ||
| for s in range(n_steps): | ||
| step += 1 | ||
| print(f"Step {step}") |
There was a problem hiding this comment.
Instead of printing steps, please use a progress bar like tqdm
|
@shalin-j Just letting you know I plan to merge this in the next day or two! |
|
@piercehowell sounds good. I'll commit the updated collision reward tonight to handle obstacles. I also have a different version I'm working on locally to separate some of the expert logic from the scenario logic, but I think we would want to discuss how to structure that at some point to ensure everything works correctly with the expert integration in rlpd, so holding off on committing those changes to this branch. |
…gentSimulator into RezasExpert
4 participants
No description provided.