BC #02: MMOneSphere, SVD pointwise, check 4DoF demo, varying ntrain + some debugging
Now checking SVD pointwise on the 4DoF demonstrator! Need to also carefully check the demonstrator since I had some bugs in my earlier code with rotations.
Created on May 2|Last edited on May 2
Comment
If it does at all good here, I would be very surprised. :)
Remember, we are still doing SVD pointwise, to give us (R,t) for action, and we clear out the x and z components when actually commanding the tool in simulation (since the real data has nothing in the x and z components, that's why it's 4DoF instead of 6DoF).
Compare with Naive method: https://wandb.ai/mooey5775/mixed_media/reports/BC-02-MMOneSphere-ntrain100-PNet2_eepose-4DoF-demonstrator---VmlldzoxOTEzODQx
Overall:
- 100 Demos: (in progress.)
- 500 Demos (good version): actually showing high success rates of 30-40% and that's very good given how rotations can cause a lot of behavioral issues... 0.3201 +/- 0.1716 for the 3X seeds after standard wandb smoothing.
- 500 Demos (buggy version): bad, but this is because rotations were going in the opposite direction. I need to understand why as this is confusing me in the code.
Success Rates, Train/Eval MSEs
All runs (ignore for plot?)
9
100 Demos (3x seeds)
3
500 Demos (3x seeds)
3
Buggy 500 Demo w/no y flip (3x seeds)
3
Example GIFs:
100 Demos
500 Demos
After 250 epochs, 1st seed, I really like the one to the bottom row, 2nd from left! That's really good and basically exactly how I implemented it.
After 250 epochs, 2nd seed, also bottom row, 2nd from left did fairly well. Again I think we should change the demonstrator, it should be a function of the object state not necessarily time (i.e., I spend a fixed amount of time on the rotation).
After 250 epochs, 3rd seed:
Other Stuff
TODO
Add a comment