Sequential Manipulation Planning for Over-actuated UAMs

Yao Su, Jiarui Li, Ziyuan Jiao, Meng Wang, Chi Chu, Hang Li, Yixin Zhu, Hangxin Liu

Introduction

Unmanned Aerial Manipulators (UAMs)—aerial platforms equipped with robotic arms—offer significant potential for interacting with environments in three-dimensional space. However, most current UAMs are limited to simple, single-step tasks. Performing complex, sequential manipulation, such as opening a drawer to store an object, remains a significant challenge. This difficulty stems from the "floating base" nature of drones; unlike ground-based robots, any movement of the manipulator exerts a reaction force that destabilizes the aerial platform. Furthermore, traditional planning frameworks often treat the drone and the arm as separate systems, leading to uncoordinated and jerky motions.

System Architecture and Hardware Innovation

To address these challenges, researchers developed a specialized over-actuated UAM platform. Unlike standard quadcopters, which are under-actuated and must tilt to move horizontally, this custom design utilizes four modular quadcopters mounted on gimbals.

Detailed hardware architecture of the over-actuated UAM, showing the modular thrust generators and the 3-DoF manipulator.

This over-actuated configuration allows the vehicle to decouple its attitude (tilt) from its translational position. By maintaining a level body while moving, the UAM provides a much more stable base for the robotic arm, enabling it to reach into confined spaces like cabinets without the risk of the body colliding with the structure due to necessary tilting. The hardware integrates a Raspberry Pi Zero 2W for communication and Dynamixel actuators for the 3-DoF manipulator.

Unified Planning via Virtual Kinematic Chains

The core theoretical innovation of this work is the Virtual Kinematic Chain (VKC) framework. In this approach, the drone’s six-degree-of-freedom (6-DoF) flight is mathematically represented as a series of virtual joints—three prismatic (sliding) and three revolute (rotating).

When the UAM interacts with an object, such as a hinged door, the object's kinematics are "welded" into this virtual chain. This unification allows a single optimization algorithm to plan the entire sequence—approaching, grasping, and manipulating—as one continuous, fluid motion. This method effectively bridges the gap between aerial robotics and the established field of ground-based mobile manipulation.

Hierarchical Control Strategy

The system employs a bi-level control architecture to manage the complexity of real-time flight and manipulation:

  1. High-Level Controller (100Hz): Running on a remote PC, this layer handles trajectory planning and feedback linearization to coordinate the whole-body motion.
  2. Low-Level Onboard Controller (500Hz): This layer manages the individual rotors and gimbal angles. A key feature is "downwash-aware" control, which accounts for aerodynamic interference between the multiple rotors to ensure stability during delicate maneuvers.

Experimental Results

The framework was validated through both high-fidelity simulations and real-world physical experiments. In simulation, the UAM successfully performed a six-step "cabinet relocation" task and a precision "light bulb installation" task.

Physical experiments demonstrating the UAM successfully opening a cabinet (top) and a drawer (bottom) to perform sequential tasks.

In physical trials, the 1.21 kg platform, tracked by a Vicon motion capture system, demonstrated the ability to fly to a desk, open a drawer, deposit a payload, and close the drawer. The hierarchical controller proved robust, maintaining trajectory accuracy even when handling payloads up to 3 kg, significantly exceeding the platform's own weight.

Conclusion and Future Directions

This research demonstrates that by combining over-actuated hardware with a unified kinematic planning framework, UAMs can achieve the level of sophisticated manipulation previously reserved for ground robots. While the current system relies on off-board computing and external motion capture, future work aims to transition these "brains" onto the drone itself. Integrating onboard SLAM (Simultaneous Localization and Mapping) and tactile sensors will be crucial for deploying these versatile robots in unstructured, real-world environments.

Summary updated

BibTeX
@misc{su2023sequential,
  author = {Yao Su and Jiarui Li and Ziyuan Jiao and Meng Wang and Chi Chu and Hang Li and Yixin Zhu and Hangxin Liu},
  title = {{Sequential Manipulation Planning for Over-actuated UAMs}},
  year = {2023},
  eprint = {2306.14105},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.14105},
}
Figure