Payload Swing Estimation and Damping Without Payload Parameters for Multirotor UAVs

What the math gives to ML

The paper offers a parameter-free disturbance-observer mechanism: infer a periodic disturbance and its unknown frequency from inertial measurements and control inputs, then inject an anti-phase correction that dissipates disturbance energy. This transfers naturally to neural-network optimization when minibatch gradients contain oscillatory components caused by momentum, alternating data structure, delayed feedback, or recurrent dynamics. The most useful implementation is an adaptive latent harmonic oscillator driven by the gradient residual, combined with a phase-aware correction to the optimizer update. Its value is falsifiable: after frequency locking, the targeted gradient or loss oscillation should decay geometrically, while instability should occur when the effective update poles leave the unit disk.