[ORBSLAM2 Agent bug] estimatedGoalPos outside map bounds

[mcimpoi]

Thank you for the great work and releasing the simulator and baselines.

I am trying to run the SLAM baseline on the MP3D val split, which runs for the first 19 episodes, and crashes on the 20th.

Steps to reproduce

1. create point_nav_mp3d_val.yaml from point_nav_mp3d.yaml provided, with:
   SENSORS: ['RGB_SENSOR', 'DEPTH_SENSOR'] and SPLIT: val
2. python habitat_baselines/agents/slam_agents.py --task-config configs/tasks/pointnav_mp3d_val.yaml

Observed Results

• when evaluating episode 20 (scene: 2azQ1b91cZZ.glb, start_position: [16.921436309814453, 0.12711000442504883, 9.845909118652344], goal_position: [4.39768648147583, 0.12711000442504883, -7.979780673980713] ), the
  IndexError: index 418 is out of bounds for dimension 3 with size 400
  (caused by: self.estimatedGoalPos2D = tensor([[201., 418.]], which is then used to set ones in the goal_map which is of size [1,1,400,400]).

Line numbers in the stacktrace are off, due to local minor edits, not changing the functionality (e.g. debug code to isolate the problem).

Traceback (most recent call last):
  File "habitat_baselines/agents/slam_agents.py", line 635, in <module>
    main()
  File "habitat_baselines/agents/slam_agents.py", line 629, in main
    metrics = benchmark.evaluate(agent)
  File "...habitat-api/habitat/core/benchmark.py", line 64, in evaluate
    action = agent.act(observations)
  File "habitat_baselines/agents/slam_agents.py", line 340, in act
    self.planned2Dpath, self.planned_waypoints = self.plan_path()
  File "habitat_baselines/agents/slam_agents.py", line 472, in plan_path
    ] = 1.0
IndexError: index 418 is out of bounds for dimension 3 with size 400

Expected:

• handle value out of bounds / not get an estimated position outside the bounds?

Relevant Code

habitat-api/habitat_baselines/agents/slam_agents.py -- plan_path()

Any hints how to fix this?
Many thanks in advance.

[mathfac]

Thank you for reaching out. cc @ducha-aiki as author of the code and @abhiskk successfully run slam baseline on val split.

[ducha-aiki]

Hi,
This is because map size for planner is limited for limiting GPU memory usage. I tried to use constant which is enough to for maps tested, yet not too big.
You can increase it here:

https://github.com/facebookresearch/habitat-api/blob/60653f7b480db604fe05786c5909a308beea9f5e/habitat_baselines/config/default.py#L40

[mcimpoi]

Thank you! 45 seems to work fine.