ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

TU Delft Professor Alonso-Mora: The Core Challenge of Drone Operation is the Coupling of Flight Stability and Contact Force

At the IROS 2026 workshop, Javier Alonso-Mora, a professor at Delft University of Technology (Netherlands), explained the essential difference between aerial and ground-based manipulation: the lack of fixed support means that contact forces generated by robotic arms directly destabilize the flying platform. His team addresses this by using hierarchical control—an outer reinforcement learning policy generates manipulation commands while an inner adaptive controller maintains stability—thereby decoupling these two objectives. This approach has been extended to multi-drone cooperative lifting. The author explicitly acknowledges current system limitations, including reliance on motion capture systems, unhandled obstacles, and assumptions of taut cables.

2026-09-30 14:29 GMT+8

When ground robots perform manipulation, there is an implicit premise: the chassis provides fixed support. When a robotic arm pushes a box or tightens a bolt, the reaction force is absorbed by the ground, leaving the robot's own attitude unaffected. Once the same robotic arm is mounted on a drone, however, this premise disappears. Forces generated by the robotic arm's movement or contact with the environment directly alter the drone's attitude and position. Flight stability and manipulation tasks are no longer two independent problems but coupled aspects of the same problem.

Javier Alonso-Mora, a professor at Delft University of Technology (Netherlands), systematically introduced how his team addresses this constraint during the "Tightly Coupled Physical Interaction and Collaboration in Multi-Robot Systems" workshop at IROS 2026.

The Autonomous Multi-Robots Laboratory led by Alonso-Mora uses a platform called the Differential Shoulder Aerial Manipulation Platform (DSAM). Its core design strategy decomposes the coupling problem into two layers: an outer layer where a reinforcement learning policy generates manipulation commands, and an inner layer where an adaptive controller handles stability control. Together, they complete the collaborative operation of the drone and robotic arm. In demonstrations, this drone completed three types of contact operations: pushing boxes, writing on a whiteboard, and tightening bolts.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Extending from single-unit to multi-unit operations makes the coupling problem more complex. Initially, Alonso-Mora's team used centralized Nonlinear Model Predictive Control (NMPC) to have three drones jointly lift an object via ropes: a central computer uniformly calculated control inputs for all drones. The system could perform trajectory tracking, resist external disturbances, and readjust after the lifted object was impacted. They subsequently attempted to remove this central control unit, switching to multi-agent reinforcement learning. During training, an Actor-Critic architecture was used, where the Critic utilized full state information in simulation to assist policy learning. During actual execution, the Critic was no longer needed; each drone ran only its own Actor network. A high-speed INDI controller operated at the bottom layer to ensure flight stability, with neural networks not directly controlling the lowest-level flight attitudes. In figure-eight trajectory tracking experiments, this distributed strategy performed comparably to centralized NMPC. Even when adding a fourth drone or causing one drone to malfunction, move in the wrong direction, or drop, the strategy demonstrated certain robustness.

However, Alonso-Mora explicitly pointed out current limitations during his presentation. While the centralized scheme could already handle obstacles, the distributed learning scheme had not yet incorporated obstacle avoidance; adding obstacles would significantly increase training complexity and time. The entire system currently still relies on motion capture equipment for positioning and has not achieved fully onboard perception. In multi-drone lifting, the model assumes cables remain taut at all times—this assumption is easily met with three drones and works reasonably well with four, but as the number increases further, some cables may go slack, causing system complexity to rise sharply. When audience members asked how multi-machine coordination could implicitly form, he directly answered, "I'm not sure either," admitting this is a common phenomenon in neural networks where it works but cannot be fully explained.

These limitations are precisely what the HARPA project aims to address. Running from 2026 to 2030, this four-year project aims to enable heterogeneous aerial robot teams to participate in construction work: some drones transport building materials, while others equipped with robotic arms handle module connection and bolt installation. Alonso-Mora stated that the project will continue to advance onboard perception capabilities, reducing reliance on motion capture. From ground to air, from single to multiple units, and from simulation to real-world construction environments, every step increases the complexity of physical constraints.

Sources

  • leiphone.com↗
Continue researching →