Rebuttal: RGBD Results (and Segm Depth)
Now we have RGBD input (woo hoo) and can test them. This report has them for all three versions (pouring, and 4d and 6d scooping). Note: RGBD for here, not RGB (unless strictly for comparison purposes). UPDATE: actually now I feel like Depth-Segmentation will give better comparisons.
Created on August 18|Last edited on August 25
Comment
Results (Scooping)
Really interesting! RGBD does seem to be doing pretty well. However I guess I am not sure I would expect TFN to be that much better if at all compared to RGBD?
The one exception is 4DoF scooping w/out augmentation, not sure why that's so bad? Though that is consistent with what we see for the RGB results ...
Update: re-doing but after https://github.com/Xingyu-Lin/softagent_rpad/pull/3/commits/c3affacb332ef34a80bed3ed75e8ce41cc19a92a
for 6DoF scooping, actually I might need to do the same for 4DoF?
Update: whew, good news, so for both 4D and 6D scooping, doing that extra correction doesn't actually do that much (most likely because rotations are predicted to be very small?). Actually I wonder if that affects any other results we might have?
Scoop 4DoF (w/augm)
5
Scoop 4DoF (no augm)
5
Scoop 4DoF (w/augm), NO svd@runtime
5
Scoop 4DoF (no augm), NO svd@runtime
5
Scoop 6DoF (no augm)
5
Scoop 6DoF (w/augm)
5
Scoop 6DoF (no augm), NO svd@runtime
5
Scoop 6DoF (w/augm), NO svd@runtime
5
Depth-Segmentation (vs RGBD)
Let's now use this for depth-segmentation and we can compare with other approaches.
Scoop 4DoF (no augm)
5
Scoop 4DoF (SEGM-DEPTH)
2
Scoop 6DoF (no augm), NO svd@runtime
5
Scoop 6DoF (SEGM-DEPTH)
5
Scoop 4DoF (no augm)
5
Scoop 6DoF (no augm), NO svd@runtime
5
Scoop 6DoF (SEGM-DEPTH)
5
Scoop 4DoF (SEGM-DEPTH)
0
Results (Pouring)
3D Pouring RGBD (no augm, incorrect depth)
5
3D Pouring RGBD (w/augm), incorrect depth
5
3D Pouring RGBD (no augm, corrected depth)
5
3D Pouring RGBD (WITH augm, corrected depth)
5
6D Pouring RGB (no augm)
5
6D Pouring RGB (WITH augm)
5
GIF Examples:
Add a comment