Meam 517 Trial Comparison

Toe XYZ Experiments

These are the best trials for the toe xyz output parameterization. In this setup, we use a gradient weighted L1 loss and a multiple output network that outputs the entire trajectory for a single decision variable at each head.

Below are the results for some of the highlighted trials that use a joint angle output. Its important to note that the loss function also changes over the trials. In the toe xyz configuration above, the L1 loss is weighted by gradient (which turns out to be velocity), while for the pure joint angle trials, the loss is a simple L1.

Below are the results for the feasibility classifiers. It's unclear what the strange spike at ~450 steps is, it's possible it's the weight decay kicking in, which results in a weird numerical runaway followed by a steep overfit. Irrespective, the network clearly learns to output feasibility.