Rebuttal: 6DoF Scooping, 25 Demos
Created on August 17|Last edited on August 21
Comment
See our 50 demo case here https://wandb.ai/mooey5775/mixed_media/reports/Rebuttal-6DoF-Scooping-50-Demos--VmlldzoyNDg4OTY0
There are a lot of different experiments so maybe we split this into results for comparing fundamentally different methods, and results for comparing ablations?
Baselines (Different Methods)Ablation Studies for ToolFlowNet Any GIFs?Baselines ...ToolFlowNet...ToolFlowNet (No Skip)
Baselines (Different Methods)
Edit: after https://github.com/Xingyu-Lin/softagent_rpad/pull/3/commits/c3affacb332ef34a80bed3ed75e8ce41cc19a92a
I think I need to re-run some of the baselines.
Edit: OK interestingly after the fix where I don't do the SVD correction, somehow that actually slightly hurt performance on average for the image-based methods (CNN). Note: RGBD results are in the other wandb
5
5
5
5
Image CNN (WITH augm)
5
5
5
5
Image CNN (WITH augm) NO svd @runtime
5
Ablation Studies for ToolFlowNet
Let's just use this for ablations, so this should be only for ToolFlowNet stuff!
TFN (lambda 0.0)
5
TFN (lambda 0.1)
5
TFN (lambda 0.5)
5
TFN (NO SKIPs)
5
TFN MSE after SVD
5
TFN PM before SVD
5
TFN (lambda 1.0)
5
Any GIFs?
Baselines ...
TODO ...
ToolFlowNet...
Epoch 500 from BC04_MMOneSphere_v02_ntrain_0025_PCL_PNet2_svd_pointwise_ee2flow_6DoF_ar_8_hor_100_scalePCL_noScaleTarg_2022_08_17_00_09_29_0005 with 100% success rate :)
ToolFlowNet (No Skip)
Yes, we are seeing the same thing, no rotations predicted.
Add a comment