Closed dosssman closed 1 year ago
Hi @dosssman, yes, you can use this to generate the oracle trajectory for either AudioGoal or Semantic AudioGoal, although it uses information that is not accessible by a real navigation agent.
Hello again.
Thank you very much for taking the time to check on this issues.
Just leaving this here for other people that might stumble upon this issue:
When I was trying to use that class, it would fail instantiating as done in the ppo_trainer.py
, it would throw and error along the lines of DummySimulator does not have '_path_finder' attribute
.
Workaround to get it working somehow, was to not use pre-rendered observations, by having config.TASK_CONFIG.SIMULATOR.USE_RENDERED_OBSERVATIONS = False
.
Another alternative I found to get optimal actions for an episode was to access envs.workers[0]._env._env._sim._oracle_actions
, which returns a list of actions that lead the agent to the goal location upon execution.
This method could even be used with USE_RENDERED_OBSERVATIONS
set to True
.
This was tested with the SAVI environment, whcih uses MP3D dataset. Not sure about Replica.
Thanks again for your time. I would consider this issue closed.
Hello there.
Long time no see.
While going through the sound-spaces files, I could not help but notice that there is a
ShortestPathFollower
class here.Would it be possible to use this class to instantiate as a form of oracle to generate optimal trajectories across the scenes ? Also, is it perhaps compatible with SAVi environments ?
Thank you for your time.