Pouring, Compare # of Demos

How does performance vary as a function of demonstration count?

Created on June 3|Last edited on June 9

Comment

﻿
I've already done the 100 demo case.  See also https://wandb.ai/mooey5775/mixed_media/reports/Pouring-SVD-Methods--VmlldzoyMTAzNDU1 for results with SVD.
06/02/2022 testing the default method with 500 demos.
SVD_POINTWISE_6D_EE2FLOW_SVD = dict﻿(
    obs_type=﻿'point_cloud'﻿,
    act_type=﻿'ee2flow'﻿,
    encoder_type=﻿'pointnet_svd_pointwise_6d_flow'﻿,
    method_flow2act=﻿'svd'﻿,
    use_consistency_loss=﻿True﻿,
    lambda_consistency=﻿0.1﻿,
    scale_pcl_flow=﻿True﻿,
    scale_pcl_val=﻿250﻿,
    separate_MLPs_R_t=﻿False﻿,
)
Section 1﻿
6D Flow SVD Pointwise Consistency 500 Demos1
6D Flow SVD Pointwise Consistency 100 Demos5
﻿
﻿
﻿
Run set1
﻿
Some GIFs of 500 demos(As of 06/04/2022 I've only run one of them but it seemed to do pretty well, but still running into DAgger-issue I think or maybe something with the simulator about the collision code? Again I'm not too concerned about this because all the algorithms are subject to the same constraints, right?)
﻿
Here's some more GIFs from this same training run from earlier epochs, ~250:
﻿
At 275:
﻿
﻿

Add a comment