Mapping a room using scenescript from meta

[ievolve]

We are building an AMR (no lidar, just a ZEDx) and have been surprised how difficult generating a useful occupancy grid can be.  I was hopeful that rtabmap Grid/Sensor config set to use the depth map would work, but alas the result is not great, and depth to laserscan feels half baked as well.

The thought is that it should be possible to at least get the walls and floor correct to get a nice base map for navigation purposes.  Then let cost maps take over for obstacle avoidance and such.

Ran into this project that seems to really nail the simple "look around and tell me where the walls and floor are at".  I was wondering/hopeful that someone had been thinking along the same lines and had tried this approach?

https://www.projectaria.com/scenescript/

[matlabbe]

That is indeed impressive, though not perfect:


It looks like this MagicPlan app https://apps.apple.com/us/app/magicplan/id427424432

That could be interesting to integrate such technology to get a clean layout of the environment / clean occupancy grid (for global planning), then rely on stereo sensor outputs for obstacle avoidance and local planning.

With ZED camera, increasing the confidence of the depth values to keep only the edges can help to project only good estimated points in the occupancy grid. ZED default depth estimation tends to overestimate textureless surfaces (to have most dense possible depth image), but it comes at a cost of creating very bad estimation on those surfaces (adding obstacles in middle of a corridor for example, or having edge depth bleeding).

For parameter tuning, you may try limiting the range of the local grid and the height of the obstacles, like:

Grid/Sensor: 1
Grid/RayTracing: 1
Grid/3D: false
Grid/RangeMax: 3
Grid/NormalsSegmentation: false
Grid/MaxObstacleHeight: 0.4
Grid/MaxGroundHeight: 0.1

cheers,
Mathieu

[ievolve]

Thanks I will give those settings a try.  

In our case we have some corridors with white walls and a few doors here and there. On one hand it is understandable that this may be difficult for depth to solve this, but when you see segmentation examples like this one (https://wizart.ai/segmentation) it leaves you hoping that there are brighter days ahead when it comes to reliably map where the open floor space is at.

I'll report back on the zed findings soon.  Thanks again for the help.