BC04: Data Augm (PourWater)
Post-CoRL. 50 demo case.
Created on July 13|Last edited on July 17
Comment
I am testing data augmentation. The test example point clouds do not have any noise added. It's just for training (and a new randomly sampled noise each time).
Overview
(07/13/2022) Data augmentation does not seem to help for pouring?? At least with no data augmentation, we do see a tiny bit of improvement in the train MSE. Maybe we didn't put in enough data augmentation?
(07/17/2022) Updated some of the results with the one that has more data augmentation and which I trained for longer. Seems like it's marginally getting better (0.496 / 0.906 = 0.547 normalized performance for the Gauss 0.05 noise) but that could be due to more training time (1000 epochs instead of 500)? It's hard to tell and I'm hesitant to try and draw some conclusions from this. Maybe it's more important in real to do this?
Note: to get the normalized performance, we need to take the best BC value from this and then to divide by 0.906 for this environment (the expert performance). For reference, the no noise case has 0.432 / 0.906 = 0.477 which is what we reported in the supplementary material.
Results
Gauss 0.01
5
Translation + Gauss 0.01
5
Gauss 0.001
5
No Data Aug
5
Gauss 0.05
5
Run set
15
Add a comment