Pouring, Naive Vector Methods, data v02
Now we're going to deal with pouring as our second environment. This has a bunch of the 'naive' baselines.
Created on June 1|Last edited on June 3
Comment
Quick Comments / Experiment LogResults with naive MSE lossesResults, Classification PN++ PoliciesResults, CNN PoliciesResults, CNN Policies (with data augmentation)Results for Naive PN++ Classification, Pointwise Loss
Quick Comments / Experiment Log
Time for us to get started!
- (06/01/2022) Run naive CNN vector policy to benchmark performance.
- 5X on cluster, 100 demos, no data augmentation.
- 5X on cluster, 100 demos, with data augmentation (random crops).
- (06/01/2022) Run naive PN++ vector policy to benchmark performance.
- 5X on cluster, 100 demos.
- (06/03/2022) Naive PN++ with pointwise loss.
- 5X on cluster, 100 demos.
- See below for the new plot with this.
Code for 06/01 runs:
Results with naive MSE losses
Naive CNN Vector
5
Naive CNN Vector DataAugm
5
Naive PN++ Vector
5
Results, Classification PN++ Policies
Interesting, most of the failures seem actually kind of similar to what we see with the SVD / flow based methods. I would have thought that determining the rotation would have been harder for this naive method. In a few cases it does seem to not realize that it has to move the tool back to the neutral position.
After training 300 epochs:
Results, CNN Policies
Interesting, it seems to struggle quite a bit.
After training 300 epochs:
Results, CNN Policies (with data augmentation)
The 4th set of GIFs looks interesting, seems like in principle it should know how to do this, but maybe it was to aggressive in raising the box, etc?
After training 300 epochs:
Results for Naive PN++ Classification, Pointwise Loss
Started this 06/03/2022 after:
Naive Class PN++ Pointwise
5
Add a comment