Erik Wijmans

The mapping between VLN-R2R panorama locations and habitat-sim is non-trivial and it's likely the infs you are running into are from the mapping not being correct. This PR did all...

> The PR removed 3220 locations with this problem from the dataset. Yeah, due to reconstruction errors and artifacts, there were a considerable number of locations that simply didn't have...

> I want to ask you about the calculation of the z value translated from matterport to habitat The z values aren't translated, the whole mesh is rotated so that...

> camera_position = agent_state.sensor_states["rgb"].position z = camera_position[1] agent_state.position = [2.345, 1.5698 - z, -1.78] > Since the camera position is not available at the start, I took the z values...

Just to clarify, `snap_point` is deterministic. i.e. `snap_point(pt) == snap_point(pt)` and if `pt` was already a navigable location, then `snap_point(pt) == pt`. Also, it will simply find the closest navigable...

> when I apply continuous action space rather than discrete action space What are you referring to here? We don't have a standardized continuous action space.

It's a little hard to tell, but this doesn't seem to be a bug. The geodesic distance between those locations is infinite, i.e. there is no navigable path from blue...

Internally, we have a check to see if there is a feasible path between two points (just checking to see if two points are part of the same connect component)....

Habitat-lab uses habitat-sim in all nearly all cases (the only case that it doesn't is when using pyrobot to deploy in reality).

Habitat assumes GLB meshes are +Z up. If yours aren't that would be an issue. Also make sure you generate a navmesh for the mesh. If you didn't, you can...