BC03: Pouring (noise to PCL)

Experiment with adding noise to point cloud.
Created on June 6|Last edited on June 15
Comment
(06/06/2022) Running experiments where we are adding noise to the point clouds. 5X should be running now. After SoftAgent code commit.
# 6D flow, but using SVD to extract actions, learning from >= 4DoF demos.
# As of 06/01/2022 this is serving as our 'proposed' method, WITH the
# consistency loss and WITHOUT the separate MLPs for (R,t).
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,
    gaussian_noise_PCL=0.001,  # normally 0
)
Should inspect the point cloud visualizations to see how much noise there is added.
Compare with no noise for the proposed method for pouring  https://wandb.ai/mooey5775/mixed_media/reports/Pouring-SVD-Methods--VmlldzoyMTAzNDU1﻿
Update (06/07) looks like this helped performance! It really suggests that we should just train with some noise. Also the curves have not yet stabilized so there's a chance we should have just done 400 epochs as a default, instead of 300?
Update: 06/13 now doing it for 3d flow as well.
Update 06/15 now doing 0.010 since 0.005 is looking surprisingly good.
Results﻿
 
6D Flow, SVD, Pointwise (Noise 0.001)5
 
3D Flow, SVD Pointwise (Noise 0.001)2
3D Flow, SVD Pointwise (Noise 0.005)5
3D Flow, SVD Pointwise (Noise 0.010)5
﻿
﻿
Add a comment