Collaborative transportation and manipulation of cable-suspended loads by multiple UAVs offer a promising way for expanding UAVs’ role in heavy-lifting operations. Existing approaches for collaborative aerial manipulation of a payload along a reference trajectory typically rely either on centralized control architectures or on reliable inter-agent communication. In this work, we propose a novel machine learning–based method for decentralized kinodynamic planning that operates effectively under partial observability and without inter-agent communication. Our method leverages imitation learning to train a decentralized homogenous student policy for each UAV by imitating a centralized kinodynamic motion planner which has access to privileged global observations. The student policy uses physics-informed neural networks that respect the derivative relationships of motion to generate trajectory that are step-wise consistent and guaranteed to be kinematically feasible. During training, the student policies utilize the full trajectory generated by the teacher policy, leading to improved sample efficiency. Therefore, the student policy can be trained in under two hours on modest hardware. We validate our method in both simulation and real-world environments to follow an agile reference trajectory, demonstrating performance comparable to that of centralized approaches.

Aerial manipulation control is challenging due to the inherently coupled and highly variable dynamics involved in simultaneously controlling both the drone platform and its manipulator. Traditional model-based methods have been widely used but are limited by their dependence on accurate dynamics models, high computational cost, and sensitivity to external disturbances. Recent studies have explored reinforcement learning (RL)-based methods to overcome these limitations. However, most of them focus on fully actuated drones or rigid-link manipulators, thereby avoiding the complexity of underactuated platforms and manipulator control. This work develops a robust RL-based controller framework capable of whole-body control of an underactuated aerial manipulator with a two-degree-of-freedom arm.
The method is evaluated in both simulation and real-world experiments on three tasks: end-effector pose control, payload carrying, and path following. We achieve 6-DoF end-effector pose tracking with errors of 5.3 cm and 8.8°, with inference times of 0.18 ms, and demonstrate payload carrying capacity of up to 140 g through real-world experiments.

This paper presents the first decentralized method to enable real-world 6-DoF manipulation of a cable-suspended load using a team of Micro-Aerial Vehicles (MAVs). Our method leverages multi-agent reinforcement learning (MARL) to train an outer-loop control policy for each MAV. Unlike state-of-the-art controllers that utilize a centralized scheme, our policy does not require global states, inter-MAV communications, nor neighboring MAV information. Instead, agents communicate implicitly through load pose observations alone, which enables high scalability and flexibility. It also significantly reduces computing costs during inference time, enabling onboard deployment of the policy. In addition, we introduce a new action space design for the MAVs using linear acceleration and body rates. This choice, combined with a robust low-level controller, enables reliable sim-to-real transfer despite significant uncertainties caused by cable tension during dynamic 3D motion. We validate our method in various real-world experiments, including full-pose control under load model uncertainties, showing setpoint tracking performance comparable to the state-of-the-art centralized method. We also demonstrate cooperation amongst agents with heterogeneous control policies, and robustness to the complete in-flight loss of one MAV. Videos of experiments: https://autonomousrobots.nl/paper_websites/aerial-manipulation-marl

Autonomous drone racing presents a unique challenge that requires both high-speed motion planning and strategic decision-making in a multi-agent setting. Prior work has primarily relied on model predictive control (MPC) methods that treat opponents as dynamic obstacles, limiting their ability to model strategic interactions. In this work, we formulate drone racing as a dynamic game and introduce game-theoretic planning methods that compute open-loop Nash equilibria, incorporate blocking strategies, and accelerate decision-making using learning-based techniques. These methods explicitly model opponent behavior, allowing drones to anticipate and react strategically in high-speed racing scenarios. To assess the effectiveness of our approach, we conduct a large-scale head-to-head tournament against MPC-based planners, demonstrating that interaction-aware planning enables more effective overtaking and defensive strategies, leading to a higher wining rate. However, computational delays in high-speed decision-making can limit performance, highlighting the need for efficient techniques that balance real-time feasibility with strategic adaptability. Our results show that learning-based acceleration significantly improves decision-making speed while preserving competitive advantages. Finally, high-fidelity simulations and real-world drone racing experiments validate the feasibility of these methods, confirming their ability to generate reliable and competitive strategies under practical racing conditions.