RMA not work as expected

[jc-bao]

After the adaptation, the training procedure is fairly slow.

[jc-bao]

Sanity check1: make sure all parameter is observable given the past trajectories.

Oracle: directly calculate parameter from past experience ✅

[jc-bao]

Sanity check2: remove useless prediction parameters and calculate the result again.

Test case: only keep mass and relative input. Need to learn $m = \frac{F_u}{a+g}$ Result: the loss went down (from 0.15 -> 0.015), but the error is still relatively high.

[jc-bao]

current vector

• How to learn the desired function which maps the trajectory to parameters.

[jc-bao]

• update loss function
• SysID paper.
• Sensitivity analysis.

[jc-bao]

Progress1: with only one parameter m

• The formula is learned with an MSE error of 0.05.
• However, when adding another parameter in, then the loss goes up to 0.9

[jc-bao]

Trail: compress the input with 3D adaptor

• Fail: MLP's performance is better than 3D adaptor

[jc-bao]

Trail: permute before reshaping

Problem resolved!

[jc-bao]

Result