MMMultiSphere (v02): BC_PCL_PNet2_avg_ee
Now doing the multi-sphere env with 1 distractor (might be hard). Here we have the segmentation PN++ with an averaging layer so we backprop through the average.
Created on April 21|Last edited on April 25
Comment
(0.4024+ 0.269 + 0.232) / 3 = 0.3011
Train/Eval MSEs, Episode Success/Reward
Run set
3
Example GIFs
First seed after 250. OH ... it seems to be getting confused between balls, interesting. They have different labels in the segmented point clouds, though...
Second seed after 250. This actually gets 0% performance, because it seems to be picking the red ball?? I wonder if this is a function of there not being that many points in the point cloud? Interestingly that could be something we might need to innovate on, if there's something more we can do for object-centric relationships between the ladle and the points?
Third seed after 250. Once again, it's getting the red ball?? Oh, actually if the yellow ball isn't visible, it's not going to be accounted for in any points, so maybe the network might be getting to that red one if the one-hot label isn't a strong enough signal?
Add a comment