🐛 Lower-Level Model Mismatch - Imprecise Attitude Rate Controller leads to MPPI unstable behavior

[jc-bao]

🚨 Problem Description:

The lower-level model-mismatch issue is affecting the performance of the MPC (Model Predictive Control). The MPC performance is suboptimal and exhibits unstable behavior, primarily attributed to an imprecise attitude rate controller.

[jc-bao]

Attempt 1: check the hovering behavior

MPPI action compare with PID action with H=1, N=8192:

MPPI action compare with PID action with H=32, N=8192: -> too noisy

try to make lamba (entroypy) larger 1e-2 -> 1e-3 -> not work try to make sampling sigma smaller -> still noisy, but with smaller scale. try to reduce horizon to 16 -> still noisy reduce H to 8 -> still noisy

sensor noise?

-> not very noisy.

MPPI behavior in simulator?

-> the same, super noisy.

💭 It might be MPPI issue, which yields high frequency actions.

[jc-bao]

Attempt: add low-pass filter to MPPI action output:

set action=0.1 action + 0.9 old action:

-> still too noisy for real-world experinment

0.05 low-pass filter:

-> still too noisy for hardware

disable model first-order delay:

hardware diverged. (I think it is caused over filtered system. )

[jc-bao]

Attempt: add quaternion cost

constrain quaternion make the system much more stable in this case. But the model mismatch still exist.

[jc-bao]

Attempt: matching hovering thrust

• hovering behavior in simulator: