IEEE Robotics & Automation Magazine - December 2018 - 82

equipped with a 2-DoF robotic arm. We use DJI E800 motors
controlled by 620S ESCs. For onboard computation, we use
Intel NUC7i7BNH running Ubuntu 14.04 and root operating
system Indigo. Also, a Pixhawk autopilot is attached at the
center of the fuselage and connected to the onboard computer
via universal serial bus (USB).
We implement a trajectory-tracking controller proposed
in [5] as a high-level controller to generate the desired net
thrust and angular velocities. These set-points are sent to the
autopilot, and we customize the PX4 firmware to compute the
desired PWM signal of each motor by a low-level control
algorithm [11], which tracks the desired angular velocities
(see Figure 6). For indoor tests, we use a motion capture system (Vicon). The GCS receives pose measurements at 100 Hz
and sends them to the onboard computer via wireless communication. Lastly, MX-106 and MX-28 servos from Robotis
constitute the 2-DoF manipulator and are connected to the
onboard computer via USB.
A red-green-blue depth or stereo camera can be used for
detecting obstacles and updating environmental information,
such as the existence or location of the obstacle. Trajectory
replanning is then accomplished by changing the environmental parameters of the current state, taking into account
the updated information.
Control Architecture
The high-level controller is meant to control the position and
heading of each aerial manipulator in a decentralized manner
[5]. We note that the high-level controller used in this experiment is a slightly modified version of that in [5], where the
desired torque output is replaced by the desired angular rates.

This is because the low-level controller implemented inside
the PX4 runs at a much higher frequency (nearly 400 Hz) and
significantly improves flight quality.
When the low-level controller receives the desired angular
rates and net thrust, it computes the desired forces to be exerted by each motor and sends the PWM signals to the ESCs.
We select a robust controller [11] as a low-level controller that
can reject disturbance by using a DOB; this controller regards
that the tracking error comes from the discrepancy between
the nominal dynamic model and the actual dynamic system.
The additional commands are generated to reduce this model
discrepancy, and the actual plant can act just like the nominal
model eventually.
Online Experiment
Based on the motion-planning framework discussed in the
"Learning-Based Motion Planning" section, we perform
experiments involving previously unknown obstacles. In the
first scenario, aerial manipulators transport an object forward while avoiding obstacles; in the second scenario, aerial
manipulators approach the waypoint, avoiding two obstacles, and then return to the initial position. At the waypoint,
the yaw angle is set to 180° opposite from the initial yaw
angle. During the experiments, each aerial manipulator
updates surrounding information when it is close to obstacles and forms a set of environmental parameters for the
new configuration. As a result, new style parameters for the
given environment are computed using the mapping function in the RRT*-PDMPs. Combining style parameters and
the basis function, the trajectory to the goal is computed
and continuously modified online.
Figure 7 shows the resulting trajectories recorded from the experiments. Figures 8 and 9 show photos
from the experiments. The time histories of each aerial manipulator are
shown in Figures 10 and 11, which
indicate satisfactory tracking of the
planned safe path.

Onboard
Raw Position
Measurements
GCS
Desired
Raw Position States
Measurements

High-Level
Controller

Desired Joint
Velocity
Manipulator

Joint
States
Desired Thrust and
IMU Angular Rates
Measurements

Environment

Low-Level
Controller

PWMs

ESCs

Figure 6. A system overview. GSC: ground station control; ESCs: electronic speed
controllers.

IEEE ROBOTICS & AUTOMATION MAGAZINE

december 2018

Conclusions
In this article, we focused on motion
planning for cooperative aerial manipulators, a topic that has not been
addressed in depth despite its importance. Because most existing motionplanning algorithms are not appropriate
for aerial cooperation in terms of
simultaneously satisfying both optimality and agility, we introduced a learning-based motion-planning algorithm.
By learning the relationship between
known environments and corresponding optimal motions with the RRT*
using PDMPs, our algorithm produced adequate trajectories for aerial

IEEE Robotics & Automation Magazine - December 2018

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2018