Adaptive Impedance Control with Payload-Mass Estimation for Aerial Catching

MuJoCo simulation of a 2R arm catching a thrown ball with compliant settling
A lateral catch in MuJoCo with the controller I have been exploring. The arm yields on impact, settles compliantly, and the base stays controllable through the transient.
A 2R planar arm on a hovering quadrotor catches a thrown ball. The controller estimates the payload mass during the impact transient, softens the impedance once a payload is detected, and keeps the wrench transmitted back to the base inside what an attitude controller can absorb. Sixteen-controller benchmark, like-for-like comparison.

The scenario I keep returning to

Picture a 2R planar arm mounted on a hovering quadrotor. Someone throws a ball into its workspace. The arm has to catch the ball, settle the contact, and do all of that without transmitting a wrench to the base that the attitude controller cannot absorb. Most aerial-manipulation work handles one of those three things well. This thread is my attempt to look at all three together, because in practice they are not separable.

What the controller is trying to do

Three ideas sit on top of standard impedance control:

  1. Payload-mass estimation runs in the milliseconds right after impact. The estimate enters the impedance law directly, so the usual stiff-to-soft tradeoff softens once the arm knows what it is holding.
  2. Like-for-like comparison. The numbers I am willing to publish are the ones that hold up against PD, sliding-mode, and naive-adaptive baselines on identical catching trajectories. The comparison is the experiment, not an afterthought.
  3. A base-wrench audit. Tip-tracking error gets most of the attention in aerial manipulation papers; the wrench transmitted to the base usually does not. Both are measured here, and the base wrench is treated as a first-class metric rather than a footnote.

Benchmark snapshots

A sixteen-controller benchmark on a shared catching scene, with identical impact conditions. The cuts are intentionally honest: the controller I am exploring does not win every column. It wins the ones I argue matter most for whether something like this could fly outside a lab.

Tracking and impedance error across baseline controllers
Tip tracking and impedance error across baseline controllers on the same catching scene.
Base wrench norm comparison across controllers
Base wrench norm, the quantity most aerial-manipulation comparisons leave out.

Video

Side-by-side comparison of the proposed controller against the baselines on the same catching scene, rendered in MuJoCo. Same impact conditions, identical reference, only the controller changes between panels.

This is one of the threads I am currently exploring, and a fair illustration of where my research interests sit: contact-rich aerial manipulation, where the controller, the model, and the platform all have to be honest with each other.