Abstract
Cooperative aerial manipulation is difficult under wind, payload motion, and noisy references because multiple UAVs are coupled through a shared load. We present CALM (Cooperative Aerial Learned Manipulation), a decentralized controller for aerial manipulators carrying a basket while tracking a moving ground robot. CALM combines a parameter-shared Multi-Agent Proximal Policy Optimization (MAPPO) policy with reference prediction, per-UAV Extended-State Observers (ESOs), payload-mass estimation, and anticipatory formation-yaw feedforward. The anticipatory formation-yaw feedforward uses estimated target motion to rotate the formation before position error develops, reducing basket-yaw lag without adding net commanded force. A disturbance-conditioned critic supports centralized training while execution remains decentralized. Across held-out trajectories, CALM achieves lower mean position and attitude errors than the matched arm baseline (7.68 versus 10.95 cm and 3.3 versus 4.7°), a cable-suspended realization, and the two evaluated model-based controllers. It remains superior to or statistically tied with the arm baseline under unseen degradation of the full information path. Videos can be found at https://petreric.github.io/CALM/.
Video
Full experiment walkthrough
Nominal tracking with no disturbance, then wind, then a moving payload center of mass, shown in sequence.
Robotic arm CAD
Assembly of the compliant arm's elements and its range of motion, as currently modeled in CAD.
Three quadrotors with compliant arms cooperatively carry a shared basket while tracking a ground robot under wind and payload-shift disturbances.
Simulation Results
CALM is evaluated in simulation against four cooperative-transport baselines — a matched decentralized MARL controller retrained on the arm platform, a native cable-suspended MARL controller, a quasi-static grasp-map wrench allocator, and RotorTM — across six held-out trajectory families (boustrophedon, circle, figure-8, perimeter, contour, and chevron) and independent sweeps of wind and payload center-of-mass motion.
Trajectory tracking
Across the six trajectory families, CALM reaches a mean position error of 7.68 cm, versus 10.95 cm for the arm-platform MARL baseline and 18.85 cm for the cable-suspended baseline. Paired per-seed differences favor CALM on all six families. The representative paths above, colored by position error, show CALM tracking closely along straight coverage, circular, and figure-eight references alike; Dec-MARL (arm) keeps pace on straight-line coverage but loses accuracy through turns, while Dec-MARL (cable) exhibits substantially larger and noisier heading error throughout, consistent with a passive cable's lack of direct attitude authority over the basket.
Disturbance sweeps
Independent wind and center-of-mass sweeps isolate each disturbance source. Under wind, CALM's position error grows from 2.85 to 10.81 cm across the tested range, while the arm baseline grows from 3.67 to 13.75 cm; CALM stays lower at every wind level tested, and both degrade continuously rather than failing abruptly. At the largest level (10.0 m/s equivalent), the cable-suspended baseline reaches 12.23 cm position error with 47° median vehicle tilt and uncontrolled load attitude. The center-of-mass sweep isolates articulated load control: from 0.05 to 0.20 m deviation, CALM changes only from 3.27 to 5.14 cm, whereas the cable baseline grows from 7.27 to 54.55 cm and remains at least 11× worse in orientation throughout. Both arm-based controllers hold basket attitude within a few degrees even at the largest deviation, where offset clipping pushes them toward the same saturated-torque regime.
BibTeX
@article{calm2027,
author = {Huo, Jiaxi and Ricioppo, Petre and Lizzio, Fausto Francesco},
title = {CALM: Decentralized Control of Cooperative Aerial Manipulation with Multi-Agent Reinforcement Learning},
journal = {TBD},
year = {2027}
}