Rebuttal: 6DoF Scooping, 50 Demos

Created on August 17|Last edited on August 17
Comment
(I would later change this to 25 demos later as it seems like results are overall very good, might need to make the demonstrator more complex or we could also add a distractor...)
Quick summary: Here in the train results we see that TFN does get a tiny bit better (0.960 at best) than the Direct Vector MSE baseline (0.904 at best), but the results are very close and might be within error ranges. TFN will probably also be better with the  avg over all the history (as the other evaluation metric) so we should get that ready in code.
But look at the 25 demo one: https://wandb.ai/mooey5775/mixed_media/reports/Rebuttal-6DoF-Scooping-25-Demos--VmlldzoyNDg5MDM0﻿
I have a LOT more results there (and they show TFN can really do well even with 25 demos).
Train Results﻿
eval/flow_episode_plot
1 - 4 of 5
BC04_MMOneSphere_v02_ntrain_0050_PCL_PNet2_svd_pointwise_ee2flow_6DoF_ar_8_hor_100_scalePCL_noScaleTarg_2022_08_16_23_53_04_0005
 plotly-logomark                        
BC04_MMOneSphere_v02_ntrain_0050_PCL_PNet2_svd_pointwise_ee2flow_6DoF_ar_8_hor_100_scalePCL_noScaleTarg_2022_08_16_23_53_04_0004
 plotly-logomark                        
BC04_MMOneSphere_v02_ntrain_0050_PCL_PNet2_svd_pointwise_ee2flow_6DoF_ar_8_hor_100_scalePCL_noScaleTarg_2022_08_16_23_53_04_0003
 plotly-logomark                        
BC04_MMOneSphere_v02_ntrain_0050_PCL_PNet2_svd_pointwise_ee2flow_6DoF_ar_8_hor_100_scalePCL_noScaleTarg_2022_08_16_23_53_04_0002
 plotly-logomark                        
Step
Direct Vector MSE (scaled all target)5
ToolFlowNet (consist. 0.1)5
﻿
﻿
﻿
Run set10
﻿
﻿
Add a comment