Pouring, Dense Transformation Methods
Let's try the dense transformations methods. It seems like this will be giving slightly better performance relative to pouring?
Created on June 5|Last edited on June 5
Comment
Let's try the dense transformations. Here the synthetic point that we always add to the point cloud will be based on the rotation center (which is [env.glass_x, env.glass_y, 0]).
- (06/04/2022) Running the dense transformation method with MSE loss. This will be the same as how we do it with the scooping env, except the center is different (due to the nature of the tool). After https://github.com/Xingyu-Lin/softagent_rpad/commit/f14eceb9a70b76da83de0b9a0f3b8cc9b06e0256
- Update: we do see good performance here, and in fact this is better than the classification PN++ with vector, so there is in fact some benefit to the segmentation PN++ here!
- (06/05/2022) Running dense transformation with pointwise loss instead (from the induced transformation).
Note: we do plot the MSEs but remember that those are not comparable among vector vs flow methods, because the latter (flow methods) will technically include a bunch of 0s from all the masked out tools, and also that the rotations are in the MSE for the vector losses but in the pointwise, the rotation is implicit when we compare the flow vectors, etc.
Results
Segm PN++, Dense TF, MSE
5
Segm PN++, Dense TF, PointWise
5
GIFs, Dense MSE
See below for the 5 seeds, performance seems reasonable.
GIFs, Dense Pointwise
TODO
Add a comment