BC03 4DoF, 1 Sphere, Dense Transformation
This is the simple dense transformation we've been talking about.
Created on May 27|Last edited on May 31
Comment
Experiment notes:
- Ran 5X runs on 05/27/2022 with 6D dense transform but with axis-angle.
- This seems less principled than pointwise loss. Yufei might have nonetheless had some success with this.
- This seems like it’s exactly the same as the “naive” PN++ classification to EE pose, except we’re using the segmentation version and picking out one of the points.
- TODO: at some point, do the dense transform but with pointwise loss. Do it (at least) after this commit where we fixed a few things.
- This seems like it’s exactly the same as the “naive” PN++ classification to EE pose, except we’re (a) using the segmentation version and picking out one of the points and (b) using a pointwise loss instead of MSE (which seems better anyway).
- Update: running 5X now! 2X from Eddie, 3X from me, 05/27/2022.
For comparison reasons, I'm also adding the "naive" method of just PN++ classification to EE pose in the plots below. See the corresponding wandb report for that HERE along with other baselines.
Update: results for all these seem awful. Why?
In particular the dense transform is even worse than the naive vector based way.
Results
Dense Transform 6D, MSE loss
5
Naive Class PN++ to 6D
5
Dense Transform 6D, PW loss
5
GIFs for Dense Transformations
GOOD NEWS: there seems to be no heavy jerks / jumps (which would indicate more simulator instability than other stuff). Here are the dense transforms with MSE. Also that the policy actually seems to be learning to go down and somewhat do the right thing ... interesting.
Interesting, see above, it does some rotation but not that good...
Now 6D Dense Transform, Pointwise
It seems to be vaguely learning to lower, but seems like it's just not doing that well, interestingly. Nothing obvious that could be causing an issue (e.g., no water instability, no jerky ladle motions, etc.) which is good news from a debugging perspective.
Note: these are the 3 I ran, and Eddie ran another 2.
Add a comment