Skip to main content

Rebuttal: 6DoF Pouring, 100 Demos

Lots of results here! :)
Created on August 22|Last edited on August 22
Remember that the demonstrator has a success rate of 0.815 and that we are reporting normalized success rates. This means for the "max" metric with Direct Vector MSE, we should report 0.176 / 0.815 = 0.216 normalized performance, which is exactly what our script reports.
Actually the Direct Vector MSE training seems to have stagnated or failed (look at train and eval MSEs). I should check this further. WAIT, no that's just 1 out of 5 seeds (fortunately).




Baselines (Different Methods)

Need to consult with Eddie, there are a lot of nuances that have to do with how we handle rotations (e.g., at test time when we convert from local to global frames).

Direct Vector MSE
5
Dense TF MSE
5
Image CNN MSE
5
Image CNN (Data Aug) MSE
5


ToolFlowNet Methods


TFN lambda 0.0
5
TFN lambda 0.1
5
TFN lambda 0.5
5
TFN PM before SVD
5



TFN lambda 0.1
5
TFN lambda 0.5
5
TFN lambda 0.0
5


Example GIFs

Baselines

ToolFlowNet Ablations

ToolFlowNet