Sensor configuration

Firstly, Thank you for your great work!

Questions

1. In configs/tasks/pointnav_gibson.yaml line 5 and configs/tasks/pointnav_mp3d.yaml line 5, SENSORS for SIMULATOR is set as [‘RGB_SENSOR’]. However, in the observations returned by the simulator, there are still ‘rgb’, ‘depth’ and ‘goal’. How can we specify the sensors we want to use in simulator observations correctly?
2. Are all the instances in the gibson/mp3d train/val/test set reachable within 500 steps? (One instance means one episode_id data like this: (I am not sure whether the word instance is suitable to describe the above data example))

[erikwijmans]

A single instance of the simulator can only utilize 1 GPU at a time, however, it is able to use any GPU when built in headless mode (using the --headless flag, this is only supported with nvidia GPUs on linux).

Currently the baseline code isn’t setup to use multiple GPUs, but it shouldn’t be hard to modify. For the simulator, you can spawn different instances on different GPUs by modifying the loop here: https://github.com/facebookresearch/habitat-api/blob/master/baselines/evaluate_ppo.py#L49 to change the GPU id based on the rank of the process. For pytorch, proceed as normal.

The coordinate system in habitat-sim is -z forward, +y up.

[Jiankai-Sun]

Thank you for your quick reply!