BC #02 MMMultiSphere "Naive" Methods, Regress Straight to 3DoF (use PN++ and RGB tests)
Naive method with classification PN++, regressing to 3DoF pose, now for multi-sphere env. Update: also testing with the RGB network.
Created on May 3|Last edited on May 3
Comment
Using the RGB images -> CNN -> raw EE translation is way better than PN++ regressing to raw EE despite the latter using 500 training demos, the former using 100. Though, the CNN has about 5X more parameters than the PN++. The interesting comparison might be the CNN Performance vs PN++ if we also add in our SVD / flow stuff.
Performance
Naive PN++ method
3
Naive CNN method
3
Example GIFs
Naive pixel-based methods. Hmm ... looking quite bad but some of it is due to env difficulty ... these are after 250 epochs when performance had leveled out and was stagnating (possibly due to overfitting)?
Code
Using these commits modulo changes in launch BC script:
https://github.com/mooey5775/softgym_MM/commit/6d137139c89c7a3fe451849fa732a91040c832f8
Add a comment