Notes on Decomposing the generalization gap in imitation learning for visual robotic manipulation

What's are the factors that contribute most to the diffculty of generalization to new environment in vision-based robotic manipulation?

To answer this question, this paper characterize environmental variables as a combination of independent factors, namely the background, lighting conditions and so on.

The results are really helpful for future experiments on simu-to-real.

1. Random crop augmentation improves generalization. This rule actually used already.
2. Visual diversity from out-of-domain data dramatically improves generalization. For example, data from opening a fridge can improve performance on picking an object.

Not read so deeply.