SCIENCE ROBOTICS | RESEARCH ARTICLE

N AV I G AT I O N Copyright © 2023 The

Authors, some
A self-rotating, single-actuated UAV with extended rights reserved;
exclusive licensee
sensor field of view for autonomous navigation American Association
for the Advancement
of Science. No claim
Nan Chen, Fanze Kong, Wei Xu, Yixi Cai, Haotian Li, Dongjiao He, Youming Qin, Fu Zhang* to original U.S.
Government Works
Uncrewed aerial vehicles (UAVs) rely heavily on visual sensors to perceive obstacles and explore environments.
Current UAVs are limited in both perception capability and task efficiency because of a small sensor field of view
(FoV). One solution could be to leverage self-rotation in UAVs to extend the sensor FoV without consuming extra
power. This natural mechanism, induced by the counter-torque of the UAV motor, has rarely been exploited by
existing autonomous UAVs because of the difficulties in design and control due to highly coupled and nonlinear
dynamics and the challenges in navigation brought by the high-rate self-rotation. Here, we present powered-
flying ultra-underactuated LiDAR (light detection and ranging) sensing aerial robot (PULSAR), an agile and self-
rotating UAV whose three-dimensional position is fully controlled by actuating only one motor to obtain the
required thrust and moment. The use of a single actuator effectively reduces the energy loss in powered
flights. Consequently, PULSAR consumes 26.7% less power than the benchmarked quadrotor with the same
total propeller disk area and avionic payloads while retaining a good level of agility. Augmented by an
onboard LiDAR sensor, PULSAR can perform autonomous navigation in unknown environments and detect
both static and dynamic obstacles in panoramic views without any external instruments. We report the exper-

Downloaded from https://www.science.org on March 18, 2023

iments of PULSAR in environment exploration and multidirectional dynamic obstacle avoidance with the ex-
tended FoV via self-rotation, which could lead to increased perception capability, task efficiency, and
flight safety.

INTRODUCTION (12). Using multiple sensors, such as cameras (13, 14), stereo
Uncrewed aerial vehicles (UAVs) have played an increasingly im- cameras (15, 16), fisheye cameras (6, 7), or LiDARs (17, 18), is
portant role in a variety of real-world applications such as search another method for obtaining larger FoVs. However, these multi-
and rescue, cave survey, building mapping, and archeological explo- sensor systems lead to additional sensor cost and processing time
ration. To fulfill the task requirements in these applications, the (19). Moreover, the component weights due to the sensors and
abilities of self-localization, environment mapping, and obstacle their processing units lead to more power consumption during a
avoidance are key. These abilities are usually based on the environ- UAV flight. These problems also exist for a UAV gimbal system ex-
mental observation provided by visual sensors on board the UAV, tending sensor FoV.
passive (e.g., RGB camera and thermal camera) or active [e.g., light Instead of merely increasing the sensor quantity (e.g., multisen-
detection and ranging (LiDAR) and infrared depth camera]. An ex- sor system) or using additional actuation (e.g., gimbal system), a
isting issue for UAVs is that the small field of view (FoV) of these more natural and power-saving method for extending the sensor
sensors severely limits the UAV’s perception capability and task ef- FoV is to use the inherent motions that the UAV is already
ficiency. Although many efforts have been made to deal with the capable of. One such motion is self-rotation, which effectively
constraints in applications induced by narrow or limited FoVs extends the sensor FoV to the full 360° without sacrificing the
(1–5), a larger FoV is still a better solution that not only reduces ability to move in three-dimensional (3D) space (20). Nevertheless,
task time by observing the environment more efficiently (6) but self-rotation brings two notable challenges: One is the UAV design
also enhances the UAV safety in the wild by perceiving dynamic ob- and control that should, on the one hand, maximally use the inher-
stacles (e.g., birds) approaching from an unknown direction (7). ent UAV motion without adding extra powered actuators and, on
the other hand, cope with highly coupled and nonlinear dynamics.
Methods for extending sensor FoVs on UAVs The other challenge resides in navigation: The high-rate rotation
Because of the advantages mentioned above, many methods have causes severe motion blurs and rapid FoV change, which markedly
been proposed to extend sensor FoVs on UAVs. One method is to degrades the performance and reliability of common visual naviga-
use a sensor with a large FoV, such as a fisheye camera (8), catadi- tion systems.
optric camera (9), and 360° LiDAR (10). However, for the fisheye
camera and the catadioptric camera, obvious distortions often Self-rotating UAVs
occur and must be properly compensated. There is also a limitation To date, several self-rotating UAV designs with a varying number of
to the installation direction due to the hemispherical FoV (11). A actuators have been explored, such as two-actuator designs (20–27)
360° LiDAR has all 360° FoVs in the horizontal direction, but the and one-actuator designs (28–33). The designs in (21, 22) used two
available vertical FoV is still narrow and has rather low resolution motors to provide the lift and moment, where the lift controlled the
altitude and the moment controlled the attitude (hence horizontal
position), hence achieving full 3D position control of the UAV. The
Department of Mechanical Engineering, University of Hong Kong, Pokfulam, Hong
Kong, China.
designs in (20, 23–27) used a motor to provide the UAV lift and a
*Corresponding author. Email: [email protected] servo to control the attitude. Compared with these two-actuator

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 1 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

designs, a one-actuator design enjoys simpler mechanical structure, (GNSS)–denied environments, such as caves, urban canyons, and
reduced energy loss, and lower component weight, which all help to dense forests. However, severe motion blurs and rapid FoV
conserve power consumption during a flight. The single-servo changes due to high-rate rotation in self-rotating UAVs cause
design in (28) used a servo to steer the UAV horizontal position great challenges for common visual-based navigation. For this
and left the UAV altitude uncontrolled because of the lack of lift reason, all self-rotating UAVs (20–33) reviewed above lacked the
(unpowered flight). Likewise, the one-motor design in (29) used a ability of autonomous navigation in unknown environments. The
motor to provide the UAV lift but left the UAV attitude (and hence work in (21, 29) mainly focused on mechantronics and control
horizontal position) uncontrolled. One-motor designs achieving without considering any navigation. In (20), a dedicated onboard
full 3D position control are found in (30–33), where the lift was pro- camera with a frame rate of more than 500 Hz was used to estimate
vided by the motor thrust (30, 31) or wing aerodynamic lift (32, 33) the UAV attitude by optical flow techniques, but the other UAV
and the attitude was controlled by cyclical adjustment of motor states, such as position and velocity, were not considered. Similarly,
thrust (30, 31) or wing lift (32, 33), respectively, during each in (23, 30), an onboard infrared phototransistor or photodiode array
period of self-rotation. These designs required a particular geome- was installed to estimate the UAV heading angle by sensing infrared
try and sophisticated mass distribution to trim the system, making it sources or optical sources in the environment, respectively. The
difficult for them to carry extra payload. Moreover, the thrust or other UAV states other than heading angle still remained unsolved.
wing lift was adjusted only once at every body self-rotation, funda- Full UAV state estimations were made in (22, 24, 31–33), but they all
mentally limiting the control rate and degrading the UAV agility relied on external motion capture systems to provide the position
and control accuracy. and attitude measurements, so they could only be used in indoor
instrumented environments. Likewise, Fregene et al. (25–27) and
Self-navigation under self-rotation Win et al. (28) estimated the full UAV states by leveraging external

Downloaded from https://www.science.org on March 18, 2023

The ability of self-navigation is crucial for UAVs to perform auton- position measurements from GPS or radio frequency instruments,
omous flights in unknown and Global Navigation Satellite System limiting their use in GNSS-denied environments. Besides the

Fig. 1. Overview of powered-flying ultra-underactuated LiDAR sensing aerial robot (PULSAR). (A) PULSAR uses one actuator (i.e., a motor) for full 3D position control
and an onboard LiDAR sensor for autonomous navigation. (B) The uncompensated motor counter-torque naturally causes a self-rotation that extends the sensor hor-
izontal FoV to 360°. (C) Autonomous flights of PULSAR in an unknown wooded environment at night; the flight trajectory is indicated by the onboard blue light-emitting
diode (LED). (D) Autonomous flights in the woods in the daytime; the flight trajectory is shown as the red path.

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 2 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

intended self-rotation in (20–33), unexpected self-rotation could self-rotating UAV is able to navigate fully autonomously in
also occur in common quadrotors in case of partial rotor failures. unknown, GNSS-denied environments during both day and night
Sun et al. (34) estimated the full state of a self-rotating quadrotor (Fig. 1, C and D). Because the sweeping of the conical LiDAR FoV
using down-facing cameras (either standard frame or event induced by self-rotation is intuitively similar to the beams of radi-
cameras). Down-facing cameras ease the state estimation (by reserv- ation emitted by an astronomical pulsar, we named this UAV
ing large FoV overlaps) but completely relinquish the FoV extension “PULSAR,” which is also an acronym for powered-flying ultra-
brought by the self-rotation. Besides state estimation, existing self- underactuated LiDAR sensing aerial robot.
rotating UAVs (20–34) lacked the ability of 3D environment We demonstrated the effectiveness of PULSAR in terms of power
mapping and hence could not navigate in unknown environments, efficiency, agility, and self-navigation ability and in both indoor and
not to mention take advantage of the extended FoV. outdoor environments. In all experiments, the UAV relied fully on
its onboard sensor and computer to estimate its states and to per-
Proposed solution ceive the environment without using any external instruments. Ex-
To solve the problems in design, control, and navigation mentioned periments showed that PULSAR consumed 26.7% less power than a
above and to take the advantages of single-actuator design at the benchmarked quadrotor with the same total propeller disk area and
same time, we propose an autonomous, single-actuated, and self-ro- avionic payloads while maintaining comparable agility in terms of
tating UAV with extended sensor FoVs. The UAV (Fig. 1A and trajectory tracking errors. PULSAR was responsive to external com-
Movie 1) uses a single actuator (motor) to control its full 3D posi- mands and showed high robustness to external disturbances, such
tion. The motor average rotation speed determines the propeller as wind gusts, making it suitable for real-world operations. With a
thrust controlling the UAV altitude, and the motor acceleration small onboard computer running an entire navigation framework
profile within each revolution determines the blade pitch angle, in real time, PULSAR successfully demonstrated autonomous

Downloaded from https://www.science.org on March 18, 2023

which induces a moment controlling the UAV attitude and hence flights in a cluttered, GNSS-denied environment. Moreover, bene-
horizontal position. Unlike the cyclical thrust adjustment in (30– fiting from its extended FoV, PULSAR was capable of detecting and
33), the blade pitch angle in our UAV is adjusted once every propel- avoiding dynamic obstacles in various directions beyond the origi-
ler rotation (as opposed to body rotation) by changing the motor nal sensor FoV.
speed, leading to a high control rate that increases the overall
UAV agility and control accuracy. Meanwhile, the motor counter-
torque naturally drives the UAV body to rotate, which uses the in- RESULTS
herent motion to extend sensor FoV without adding extra powered UAV system overview
actuators (Fig. 1B). Using the minimum number of actuators, the Shown in Fig. 2A, PULSAR has an overall weight of 1.23 kg, of
proposed UAV has a simple mechanical structure and effectively which 50% (616 g) is contributed by the payload (i.e., 3D LiDAR,
reduces the energy loss and component weight caused by actuators, onboard computer, and wires). At static, the motor drives the pro-
which all help to decrease the overall power consumption. More- peller to rotate up to 6800 rpm, providing a maximum thrust of
over, the symmetric UAV structure allows easy mounting of pay- 25.45 N or a thrust-to-weight ratio of 2.1. PULSAR has a diameter
loads, such as a 3D LiDAR sensor and an onboard computer, of 37.6 cm, which is also the propeller diameter, and a height of 23.7
without changing the system controllability. By taking advantage cm. By actuating the single motor and using the onboard sensor and
of the active and high-rate LiDAR measurements, the proposed computer, PULSAR achieves full 3D position control and complete-
ly autonomous flights in unknown environments.

Mechanical design
The mechanical design of PULSAR is shown in Fig. 2A. It mainly
consists of three modules: a flight control module at the top (i.e.,
propeller, motor, and flight controller), a 3D LiDAR sensor with
an onboard computer in the middle, and a battery chassis and
landing gears at the bottom. This modular design allows the UAV
to be easily maintained and reconfigured (see fig. S9). The holding
structure for each module was 3D-printed with nylon materials,
leading to a compact and rigid structure. The counter-torque of
the motor drives the entire UAV body attached with the motor
stator to rotate in the opposite direction of the propeller rotation,
which then counteracts the propeller speed in an inertial frame
and hence reduces the propeller thrust. To mitigate this effect, we
used four anti-torque vanes to provide air drag that compensated
for the motor counter-torque and hence restricted the body rotation
rate to around 2.7 Hz. The landing gear consists of three wheels
formed into a circle such that the UAV can start rotating on the
ground before takeoff.
Movie 1. Summary of PULSAR. This video shows the LiDAR scan during self-ro-
tation, the process of control moment generation, and the real-world flights in
both indoor and outdoor environments of PULSAR.

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 3 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on March 18, 2023

Fig. 2. Mechanical structure, avionics, and swashplateless mechanism. (A) Components description. (B) Interconnection among all electronic components. (C) De-
tailed mechanical structure of the swashplateless mechanism. (D) Assembly of the swashplateless mechanism, propeller blades, and motor. (E) Process of rotor accel-
eration causing the blades to lag from the rotor. With the acceleration voltage being applied at rotor position I (i.e., angle α), the rotor accelerates and reaches rotor
position II (i.e., angle α + λ0). Because the hinges between the blades and the rotor provide an extra rotation degree of freedom, the blade inertia will cause it to rotate
along the hinge and produce a blade lag angle (i.e., δ) behind the rotor. The actual blade position after the acceleration is at position angle α + λ0 - δ, as indicated by the
yellow dashed line. (F) Because of the tilted asymmetric hinge design, a lag from the rotor will cause the red blade to increase its pitch angle and the blue blade to
decrease its pitch angle from the nominal pitch angle indicated by the purple blade. The differential pitch angle changes lead to a net moment MC within the propeller
disk plane with a direction perpendicular to the actual blade position [i.e., angle α + λ0 - δ + π/2, as indicated by the purple arrow in (E)].

Avionics a common electronic speed controller (ESC) of the model

The avionics of PULSAR are all off-the-shelf available. As shown in CYCLONE 45A. The motor command is computed by the flight
Fig. 2B, they are divided into five modules. The entire system is controller once a rotor angle measurement (by the magnetic
powered by a six-cell 41.07-Wh (1850-mAh) battery. The encoder) arrives at the pulse-width modulation (PWM) capture
onboard computer is a Khadas VIM3 Pro with an ARM processor, port. The computed motor command is lastly sent to the ESC via
which runs the core navigation modules in real time. The navigation DShot600 protocol, which enables up to 33.3-kHz command
module plans a trajectory that is sent to a flight controller (Pixhawk signal transmission.
4 Mini) for tracking. The sensing module contains a Livox AVIA
LiDAR, which measures 3D point clouds at a frequency of Swashplateless mechanism
240,000 points per second in a conical FoV of about 70° (Fig. 1B). The key component that enables the modulation of a moment on
Besides the LiDAR, a magnetic encoder (AS5600) was installed on the propeller is a swashplateless mechanism invented in (35). As
the bottom of the motor stator (see the Supplementary Materials), shown in Fig. 2 (C and D), the mechanism consists of two passive
which measures the rotor’s angular position (i.e., rotor angle) at 910 hinges connecting two side hubs to the central hub attached to the
Hz. This provides about 11 measurements per motor revolution motor stator. The two hinges are parallel; both are tilted from the
when the stator is at 5000 rpm (i.e., 83.3 r/s). The actuation motor axis by 45°. Such a specially designed mechanism is able to
module contains a motor (T-MOTOR MN5006 KV450) driven by adjust the cyclic pitch angle of the two blades clamped on the side

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 4 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

hubs, a function that has been traditionally realized by a complicat- less than that of the 8-inch propeller quadrotor (which has a much
ed and high-cost swashplate mechanism in helicopters. The change larger propeller disk area). The overall efficiency of PULSAR is 17.5
of cyclic pitch angle causes a moment in the propeller disk plane and 4.1% more than that of the 7.5- and 8-inch propeller quadrotor,
that can then control the UAV’s attitude in pitch and roll (movie respectively.
S1). The working principle of this mechanism and its driving Moreover, we compared PULSAR with three commercial quad-
method are described in more detail in Materials and Methods. rotor UAVs from DJI, including Mavic Air 2, Mavic 3, and Phantom
In our implementation, the swashplateless mechanism was 3D- 4 Pro V2.0. Because the standard PULSAR has a different weight
printed with tough polylactic acid materials, and the propeller and propeller disk area from the commercial quadrotors, to make
blades are standard T-MOTOR 13-inch MF1302 propeller blades. a fair comparison, we developed two variants of PULSAR: The
For the swashplateless mechanism to induce moment more easily, first one, denoted as “PULSAR (no LiDAR),” removes the LiDAR
the motor T-MOTOR MN5006 KV450 was chosen according to a payload, and the second one, denoted as “PULSAR (no LiDAR,
quantitative analysis detailed in the Supplementary Materials. 16.4″ propeller),” additionally scales up the propeller to 16.4
The original swashplateless design in (35, 36) uses a hinge di- inches (see fig. S13).
rectly contacting the moving parts, which brings obvious friction The comparison results are shown in Fig. 4 and table S1. As can
on the contact surface. The friction then prevents the blade pitch be seen, PULSAR (no LiDAR, 16.4″ propeller) achieved the longest
angle from responding to small motor acceleration commands, re- hover time among all UAVs under comparison and outperformed
sulting in a “deadband phenomenon” (i.e., no moment output when its counterpart DJI Mavic 3 in terms of flight time (40 min 27 s
the inputed command is lower than a threshold) in this mechanism. versus 40 min) and efficiency (8.21 g/W versus 7.75 g/W), despite
The deadband phenomenon brings nonlinearity to the system and having disadvantages in every aspect, including a battery with less
considerably degrades the overall control performance and system energy (73.26 Wh versus 77 Wh), less total disk area (1362.8 cm2

Downloaded from https://www.science.org on March 18, 2023

agility. Moreover, the friction also causes energy loss and material versus 1790.9 cm2), and more total weight (929 g versus 895 g).
wear, which lowers the power efficiency and reliability. To overcome Next, comparing PULSAR (no LiDAR) with DJI Mavic Air 2, al-
this issue, we used four ball bearings between the central hub and though the former has a considerably larger weight (750 g versus
the smooth shaft screws and two pressure bearings and two steel 570 g), it still achieved higher power efficiency (8.20 g/W versus
shims between the central hub and the side hubs. They can consid- 7.76 g/W). More comparison analyses are supplied in the Supple-
erably reduce the friction in contact surfaces, even under high-stress mentary Materials.
conditions caused by high-rate rotation of the propeller and cyclic
twisting of the asymmetric hinges. Consequently, the propeller Trajectory tracking in indoor environment
moment shows no visible deadband phenomenon (fig. S2). To evaluate power efficiency in dynamic flight conditions as well as
maneuverability, we conducted trajectory tracking experiments on
Experimental validation both PULSAR and the built quadrotor in an indoor environment. In
We performed various real-world experiments to verify the perfor- the experiment, the trajectory command was set to a figure “8” path
mance of PULSAR. In all experiments, PULSAR used the same that was planned offboard (trajectory commands switched to T3 in
LiDAR-inertial odometry and trajectory-tracking controller to esti- Fig. 3A). To fully test the UAV dynamics, as shown in Fig. 5, we
mate its full state and track the trajectory commands, respectively expanded the figure “8” path in all three directions (3.6 m in x,
(Fig. 3A). The trajectory commands were generated by the 1.2 m in y, and 1.0 m in z). The trajectory period T was adjusted
onboard trajectory planner, a dynamic obstacle detector and to change the flight speed (i.e., T = 8, 6, or 5 s): A smaller period
planner, or the preset trajectory library, depending on the experi- leads to a higher average speed. Each trajectory was executed for
ment purposes. The software framework overview is described in five cycles to characterize the average power consumption and
Materials and Methods. tracking accuracy. To make a fair comparison, both PULSAR and
the quadrotor used the same LiDAR-inertial odometry method to
Flight efficiency provide the state feedback and the same cascaded proportional-in-
To verify the flight efficiency of PULSAR, we compared its power tegral-derivative (PID) controller structure with velocity and accel-
consumption with two benchmarked quadrotors both carrying eration feedforward to track the trajectory (see details in Materials
the same avionic payloads (i.e., companion computer and LiDAR and Methods). The controller parameters of the two UAVs were
sensor): One quadrotor has 7.5-inch propellers, leading to the tuned to our best efforts.
same total disk area as PULSAR, and the other has 8-inch propel- Figure 5A shows the tracking performances of PULSAR. Overall,
lers, leading to 16.8% more disk area than PULSAR (fig. S5). For PULSAR tracked these 3D trajectories tightly, with errors below a
each benchmarked quadrotor, we selected the propulsion systems small bound. At a period T = 8 s (maximal speed, 1.93 m/s), the
with the best efficiency among six motors and nine propellers tracked trajectory matched with the reference trajectory closely, in-
(see the “The benchmarked quadrotors” section in the Supplemen- dicating a good tracking performance at low speeds. The tracking
tary Materials). performance degraded with the increase in flight speed. At a
We validated the efficiency improvement of PULSAR against the period T = 5 s (maximal speed, 3.01 m/s), the tracked trajectories
benchmarked quadrotors in actual hover flights. In the experiment, exhibited obvious deviation from the reference. This performance
the trajectory command was set to a stationary hovering position, as degradation with flight speed is a typical phenomenon for multi-
shown in movie S2 (trajectory commands switched to T3 in copter UAVs due to the increased disturbances, such as blade flap-
Fig. 3A). Results in Fig. 4 and table S1 show that the total power ping (37). Tracking high-speed trajectories also requires fast actua-
consumption of PULSAR is 26.7% less than that of the 7.5-inch pro- tor responses that are limited by the motor delay. Figure 5 (B and C)
peller quadrotor (which has a similar propeller disk area) and 17.3% shows the quantitative comparison of PULSAR and the quadrotor.

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 5 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on March 18, 2023

Fig. 3. Software framework. (A) Overall software framework. The trajectory commands were switched among different sources according to the application. Details of
the position and velocity controller are shown in fig. S12. (B) Map used by trajectory planners to represent static and dynamic obstacles in environments. The map is
named “ikd-forest,” which cuts the space into 1-m voxels, each containing an incremental k-d tree data structure, ikd-tree (57). The ikd-tree downsamples all points
contained in the voxel with 0.1-m spatial resolution by retaining the centermost point on the tree. Besides the point coordinate pi, each node of the k-dimensional
tree saves the total number of the hitting point counter ci and the last hitting point timestamp ti.

The average power consumption of PULSAR was around 186 W, Response to position commands
whereas those of the two benchmarked quadrotors are about 255 To verify the agility of PULSAR, we tested it to respond a step po-
and 226 W for 7.5- and 8-inch propellers, respectively, suggesting sition command (movie S4). In the experiment, the trajectory
better efficiency of PULSAR. The tracking accuracy was assessed by command was set to a step signal in the x direction (trajectory com-
the absolute position error, which is the norm of the position error mands switched to T3 in Fig. 3A). Figure 6A shows the step re-
between the reference trajectory and the tracked one. The results sponse. The response in the x direction suggests a rise time of
suggest that PULSAR achieved slightly larger, yet acceptable, track- 0.91 s and a settling time of 2.21 s without any noticeable steady-
ing errors than the quadrotor. Figure 5E shows the 3D maps built by state error. In responding to the command in x, the position in y was
the two UAVs during the flights. Benefiting from the extended FoV also perturbed because of the self-rotation–induced gyroscopic
via self-rotation, PULSAR had a more complete mapping of the en- effect, leading to a maximum position error of 0.173 m. Similarly,
vironment, whereas the quadrotor only mapped a small portion the tilting of thrust to respond to the command in x also caused a
because of the constant yaw angle during the flight. The process temporary altitude drop, leading to a 0.101-m maximum position
of PULSAR tracking a trajectory with a period of 5 s is shown in error in the z direction. Figure 6 (B and C) shows the attitude and
movie S3. angular velocity responses, which demonstrate tight tracking of

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 6 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on March 18, 2023

Fig. 4. Hover time, power consumption, and efficiency comparisons of PULSAR, the benchmarked quadrotors, and the commercial UAVs. The data of PULSAR
and the benchmarked quadrotor are from experiments, and the data of the commercial UAVs are from DJI official specifications.

pitch and roll commands along with the uncontrolled yaw angle and its stability and hovering position in the presence of external distur-
rate, a natural result of the self-rotation design of PULSAR. bances, making it suitable for operation in real-world
Figure 6D shows the three control actions controlling pitch, roll, environments.
and altitude, respectively. These three actions were synthesized
into a single command (Fig. 6D, iv) for the motor to execute. Autonomous navigation in unknown, GNSS-denied
environments
Robustness to external disturbances To verify the full autonomous navigation ability of PULSAR, we
When operating in outdoor environments, UAVs are usually sus- performed a waypoint navigation experiment in a wooded environ-
ceptible to certain disturbances, such as wind gusts. To validate ment of 54 m by 26 m (Figs. 1D and 8A). In the experiment, the
PULSAR’s robustness to such external disturbances, we used a fan trajectory command was computed in real time by the onboard tra-
to produce a wind gust and measured the displacement of PULSAR jectory planner (trajectory commands switched to T1 in Fig. 3A),
from its initial hovering position (movie S5). Figure 7 shows the ex- which had no prior knowledge of the environment except for
perimental setup and results. In the beginning, PULSAR was hov- eight waypoints scattered in the area (star points in Fig. 8A). As
ering at a height of 1 m and a position of 0.7 m in front of the fan the flight proceeded, the trajectory planner automatically generated
(trajectory commands switched to T3 in Fig. 3A). The fan at this a smooth trajectory (orange path in Fig. 8A) from the UAV current
distance created a wind gust with speeds up to 4.5 m/s for position to the next waypoint, without colliding with any obstacles
PULSAR. As the wind was applied, PULSAR was pushed away (e.g., trunks, branches, and tree leaves). The planned trajectory was
from the hovering position by a maximum of 15.6 cm. The position then tracked in real time by the onboard controller with small errors
error caused the flight controller to estimate and to compensate for (blue path in Fig. 8A). The total flight time was about 125.7 s over
wind disturbance, bringing PULSAR back to its original position the 63-m trajectory, during which PULSAR flew fully autonomous-
after 11.5 s. Similarly, when the fan was off, the redundant compen- ly without any human intervention or piloting. Moreover, a 3D
sation action led to a displacement up to 18.0 cm in the opposite point cloud map of the environment was obtained during the
direction. This position error caused the flight controller to adjust flight. Benefiting from the extended FoV of PULSAR, the built
the compensation action and, after 7.2 s, brought PULSAR back to map had points uniformly distributed in all horizontal directions,
the original position. During the response to wind applied on the x instead of all lying within the small conical sensor FoV, leading to a
direction, PULSAR’s position in y and z was slightly perturbed more efficient exploration of the environment. We conducted the
because of the coupling effect, causing a maximum position error experiment successfully in the same wooded environment in both
of 8.2 cm for y and 4.1 cm for z. Moreover, the pitch and roll re- day and night (see Fig. 1, C and D, and movie S6) and in other en-
sponses tightly tracked their respective desired values computed vironments, such as a cave (movie S7). The various successful flights
by the outer-loop position controllers, whereas the yaw was freely suggest that PULSAR has a robust navigation ability in unknown,
rotating. This experiment suggests that PULSAR is able to maintain GNSS-denied environments.

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 7 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on March 18, 2023

Fig. 5. Indoor trajectory tracking. (A) The trajectory under tracking is a figure “8” path with different periods, T. The green line represents the reference trajectory,
whereas the line colored by velocity represents the actual flight path. Each trajectory is tracked for five cycles, where the middle three are displayed for better visualization
of the speed. A(i), A(ii), and A(iii) are the trajectories with period T of 8, 6, and 5 s, respectively. (B) Power consumption during hover and trajectory tracking. The hover
power is filtered by a moving average filter of window size 6. (C) Absolute position error of hover and trajectory tracking. On each box plot of (B) and (C), the central mark
indicates the median, and the green “+” symbol indicates the mean. The bottom and top edges of the box represent the 25th and 75th percentiles, respectively. The
whiskers extend to the maximum and minimum values excluding the outliers, which are plotted individually using the orange points. (D) Overlaid snapshots of PULSAR
when it is tracking the trajectory with period T = 8 s. (E) 3D maps built by the two types of UAVs during trajectory tracking.

Figure 8 (B and C) shows the behavior of PULSAR in response to The computation time of the navigation modules running on the
balls from two different directions. In both cases, the images on the onboard ARM processor is shown in Fig. 8D. The average compu-
left are overlaid snapshots of PULSAR and the ball captured at suc- tation times of the LiDAR-inertial odometry and the trajectory
cessive moments. PULSAR successfully detected the ball once it planner were 9.48 and 8.42 ms, respectively, which were well
faced the ball due to self-rotation. Then, it moved to one side to below the 20-ms period (i.e., 50 Hz) and suggested a real-time per-
avoid the ball while maintaining a safe margin with other static formance of the navigation module. Note that the trajectory planner
structures (e.g., trees). The images on the right are third-person and the odometry ran in parallel (see Materials and Methods), so a
views from the global point cloud maps built by PULSAR, including real-time performance would only require each individual module
the detected ball position (gray ball markers), ball trajectory (yellow to take no more than 20 ms.
arrow), point measurements in the current LiDAR FoV, and the
complete UAV trajectory (blue path) from takeoff to hovering Avoidance of dynamic obstacles from different directions
and then to avoid the ball. These results imply that PULSAR is To demonstrate the advantage brought by the extended FoV
able to perform agile motions and to perceive the environment through self-rotation, we tested PULSAR’s ability to avoid
(both dynamic obstacles and static structures) in all horizontal di- dynamic obstacles in an outdoor environment (movie S8). We
rections beyond the sensor’s original FoV. threw a ball from two orthogonal directions that could not be de-
tected if PULSAR did not have self-rotation due to the 70° conical

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 8 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on March 18, 2023

Fig. 6. Response of PULSAR to a step position command in the x direction. (A) Position response: (i to iii) Position responses in x, y, and z, respectively, and (iv) the
position errors in x, y, and z. (B) Attitude response: (i and ii) Responses in the UAV roll and pitch angles, (iii) the UAV yaw angle, and (iv) attitude errors defined as the
intersection angle between zIB and zIB;d , which excludes the yaw angle difference. (C) Angular velocity responses: (i and ii) Angular velocity response in x and y represented
in inertial frame I , respectively, and (iii) angular velocity in z represented in body frame B, which indicates the self-rotation of the UAV body. (D) Control actions: (i) The
desired moment MBd (see Eq. 14, which has components in roll and pitch only) and the thrust fT, d (normalized by the maximal thrust fT, max; see Eq. 11). (ii) The amplitude of
the sinusoidal throttle us and the complete sinusoidal throttle modulated on the rotor angular position ϕ, us sin (ϕ - α). (iii) The average throttle ua and (iv) the total throttle
ut output to the ESC. Note that this is the only actuator command.

sensor FoV. The ball had a diameter of 32 cm, a size similar to respectively, which lead to a small detection latency and reliable
PULSAR, and approached at a speed of around 5.8 m/s. We threw 50-Hz real-time running.
the ball such that it would have hit PULSAR if no evasive maneuver
was taken. Consequently, to avoid this collision, PULSAR must
detect the incoming ball along with static obstacles in the environ- DISCUSSION
ment and generate a safe target position for execution (trajectory PULSAR’s key features
commands switched to T2 in Fig. 3A). PULSAR has a unique design compared with existing single-actua-
We conducted the experiment multiple times, and the vehicle tor UAVs. The dSAW proposed in (28) used a servo to drive the
managed to detect and avoid the ball most of the time. The major cyclic pitch of a wing flap during an air-induced self-rotation.
failure cases were caused by short triggering distances, which is the Because of the lack of propellers, the dSAW cannot perform
distance of the ball that triggers PULSAR to execute an evasive ma- powered flight and only works by dropping from a preset altitude.
neuver. When a ball arrived at the triggering distance, it might just The single-motor designs (29–33) are most similar to PULSAR in
miss the LiDAR FoV and hit the UAV body before it could be de- terms of the number of actuators. Among them, Piccoli and Yim
tected in the next revolution (after 0.3 s due to ~2.7-Hz self-rotation (29) used the motor to provide the lift only, leading to uncontrolled
rate). In practice, this drawback could be trivially overcome by in- attitude and lateral motion. The designs (30–33) achieved full 3D
creasing the triggering distance. In the experiment, the triggering position control using only one motor, similar to PULSAR, but
distance was set to 4 m because the ball cannot travel a longer dis- their flying and actuation principles were completely different. Spe-
tance in the air due to gravity. cifically, Piccoli and Yim (30) and Zhang et al. (31) installed the
The computation time of the navigation modules is shown in motor on one side of the UAV frame, a position that is displaced
Fig. 8E. The average computation times of the odometry and from the center of mass (CoM). Then, with a normal propeller at-
dynamic obstacle detector and planner were 11.35 and 1.15 ms, tached, the noncentral motor at rotation provided a thrust that, on
the one hand, lifted the UAV’s altitude and, on the other hand,

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 9 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

Downloaded from https://www.science.org on March 18, 2023

Fig. 7. Wind disturbance rejection. (A) The position error (i), attitude (pitch and roll) response (ii), and angular velocity in the z axis (iii). The position error (i) is the
difference between the trajectory command (which is the specified hovering position) and the position feedback. The angular velocity z (iii) indicates PULSAR’s self-
rotation. (B) Experiment setup. The UAV is hovering 0.7 m away from the fan and at a height of 1 m. The fan creates a wind gust of 4.5 m/s at the hovering position
of PULSAR.

produced a pitch torque that inclined the thrust from the vertical build a 3D point cloud map of the environment, from which both
direction. Meanwhile, the counter-torque of the motor caused the static and dynamic obstacles can be detected and avoided in real
UAV, hence the inclined thrust, to spin. In the case of hovering, the time. Consequently, PULSAR can safely navigate in a variety of
motor speed was kept constant; then, the inclined thrust at spinning GNSS-denied environments (e.g., woods, caves, and tunnels), an
will contribute to a net force along the vertical direction only. To ability that was rarely demonstrated in existing self-rotating UAVs.
move the UAV along a horizontal direction, the motor speed was The swashplateless mechanism used by PULSAR is not the first
increased at the moment the thrust was inclined to that direction, time that it was used on a micro UAV. Previous works (35, 36, 38)
hence actuating the lateral motion once every revolution of the UAV used the swashplateless mechanism on coaxial-rotor UAVs
self-rotation. On the other hand, Win et al. (32, 33) installed the (CRUAVs) with two propellers. Besides the apparent differences
motor on a one-side wing of the UAV. The motor provided a in the self-rotation ability and number of actuators, which were
thrust driving the wing to rotate, which produced an aerodynamic the primary motivation of this work, PULSAR also differs from
lift that controlled the UAV’s altitude. To move the UAV along a these works in many implementation details. One major improve-
horizontal direction, the motor speed (hence thrust) was increased ment in PULSAR is the decreased friction in the swashplateless
or decreased at the respective location, which drove the wing to de- mechanism, eliminating the deadband phenomenon (i.e., no
celerate (lower lift) or accelerate (higher lift). The differential aero- moment output when the input command is lower than a thresh-
dynamic lift led to a moment that actuated the UAV pitch or roll, old) in CRUAVs. Another improvement is the simplified motor
hence controlling the UAV lateral motion once every revolution of controller implementation. CRUAVs used a dedicated motor ESC
the UAV self-rotation. In contrast to (30–33), the lateral motion of to drive the swashplateless mechanism, whereas, in PULSAR, the
PULSAR is actuated at every revolution of the propeller rotation, driving strategy is implemented on the flight controller without
which is at a much higher rate than the body self-rotation (4700 any such ESC. All electronics in PULSAR, including the ESC,
rpm of the propeller versus 160 rpm of the body). The increased flight controller, LiDAR sensor, and onboard computer, are avail-
actuation rate enables more accurate trajectory tracking, rapid able off-the-shelf, making the implementation much easier.
step response, and robust disturbance rejection for PULSAR,
which were not demonstrated previously (30–33). Energy efficiency
PULSAR is also a self-rotating UAV that can navigate autono- PULSAR has a greater energy efficiency when compared with the
mously in unknown environments. By leveraging the high-rate benchmarked quadrotors. According to the momentum pffiffiffiffiffiffiffiffi ptheory
ffiffiffiffiffiffi
LiDAR point measurements, PULSAR can robustly estimate its (39, 40), a UAV has its ideal hover efficiency η ¼ 2ρA= mg (g/
full state in the presence of fast FoV changes, requiring no external W), where m, A, ρ, and g are total mass, total propeller disk area, air
instrumentation. With the LiDAR sensor, PULSAR is also able to density, and the gravity acceleration, respectively. For a quadrotor

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 10 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

Fig. 8. Outdoor experiments with LiDAR sensor. (A)

Autonomous waypoint navigation of PULSAR in an
outdoor wood environment. The start point and the
destination point are the same. The time shown on each
waypoint is the time that PULSAR arrives at that way-
point. During the flight, a 3D point cloud map of the
environment is simultaneously built. (B and C) Dynamic
ball avoidance with PULSAR. The ball was thrown from
two orthogonal directions whose avoiding processes are
shown in (B) and (C), respectively. In both cases, overlaid
snapshots on the left show the ball trajectory, the posi-
tion of the ball when it is detected by PULSAR, and the
avoiding trajectory executed by PULSAR. Images on the
right show third-person views of the environment map,
current LiDAR point measurements, the detected ball
position and its trajectory, and the complete UAV tra-
jectory before and after avoiding the ball. Poses of the
camera capturing snapshots on the left are indicated by
the green arrows. The environment map is simultane-
ously built by the LiDAR-inertial odometry system
running on the onboard computer. (D) Computation
times of the odometry and trajectory planner in the

Downloaded from https://www.science.org on March 18, 2023

autonomous waypoint navigation experiment. (E)
Computation times of the odometry and dynamic ob-
stacle detector and planner in the dynamic obstacle
avoidance experiment.

that preserves the same total disk area and weight as PULSAR, it rotor-to-rotor interactions (41–43) and rotor-to-body interactions
should have the same power consumption and efficiency. (44), which would cause a further efficiency drop (for the 7.5-
However, in practice, four small propellers often have around 5.79 inch propeller quadrotor, the drop is about 5.99%). In contrast,
to 13.61% lower efficiency than one big propeller (fig. S6), due to the PULSAR eliminates this rotor-to-rotor interaction and mitigates
different Reynolds number and propeller geometry that are not ac- the rotor-to-body interaction by placing the UAV body under the
counted for by momentum theory (39). Then, installing the four propeller hub, leading to an efficiency drop of merely 2.88%.
propellers on the quadrotor airframe would further introduce Next, a quadrotor UAV has more component weight (due to the

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 11 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

higher number of propulsion systems) and structure weight (due to than that of PULSAR (2.1 thrust-to-weight ratio), respectively.
the distributed motor location on four arms), which bring more The low thrust-to-weight ratio has prevented PULSAR from execut-
power consumption and efficiency drop. In our case, the 7.5-inch ing extremely agile maneuvers, such as flips and racing. Further in-
propeller quadrotor has 206 g more weight than PULSAR (table creasing the thrust-to-weight ratio is possible by increasing the
S3), which leads to a 4.35% efficiency drop. Considering all the propeller size or decreasing the weight; nevertheless, the former
above factors and the power of onboard avionic devices (table would prevent the UAV from flying in tight spaces, whereas the
S4), PULSAR has an overall efficiency that is 17.5% higher than latter requires more lightweight sensors suitable for navigation at
the 7.5-inch propeller quadrotor (which has a similar total disk high rotation rates.
area) and 4.07% higher than the 8-inch propeller quadrotor
(which has a larger total disk area), respectively. More detailed anal- Scalability
yses and comparison results are in the Supplementary Materials. PULSAR can be adapted to different scales because of the good scal-
Besides the custom quadrotors, PULSAR also exhibits efficiency ability of the swashplateless mechanism, which has been adequately
advantages over commercial quadrotor UAVs with smaller weights verified by experiments conducted on prototypes with diameters
and larger sizes. The smaller weights and larger sizes already place from 0.1 to 1 m (53). Other parts of PULSAR, including the body
commercial quadrotor UAVs in an advantageous position, but structure, vanes, landing gears, and battery, are either manufactured
PULSAR still achieved a higher efficiency (see the Supplementary with common materials and fabrication techniques or commercial
Materials for more details). off-the-shelf components, which all scale well. The scalability of
PULSAR along with its symmetric geometry allows for carrying
Mapping efficiency various payloads of different weights as required by the task. For
The mapping efficiency of PULSAR is mainly attributed to two instance, by adopting motion blur correction (20) or ego-motion

Downloaded from https://www.science.org on March 18, 2023

factors: First, the increased energy efficiency enables longer-time compensation (54), a mini-size high-speed RGB camera or event
tasks, avoiding frequent returns and recharges (6, 45); second, the camera can be installed for a smaller version of PULSAR, whereas
extended FoV via self-rotation measures more volumes per unit larger-scale equipment such as radar, zoom camera, or LiDAR
time by naturally distributing the sensor measurements equally camera system can be carried by a bigger version.
around the UAV without any active planning on the yaw angle.
To quantify the FoV extension by self-rotation, we used the Potential applications
concept of solid angle (fig. S10). The current LiDAR on PULSAR The high energy efficiency, autonomous navigation ability, and ex-
(Livox AVIA) has a conical FoV of about 70° in both vertical and tended sensor FoV of PULSAR make it very suitable for exploration
horizontal directions, resulting in an approximate solid angle of tasks, such as environment surveying, search and rescue, disaster
1.40 sr at static. By self-rotation, the solid angle reaches 7.21 sr relief, terrain mapping, and automatic 3D reconstruction. Environ-
(57.4% of omnidirection), which is 5.15 times higher. In the case ments in these tasks are often unknown and GNSS-denied (e.g., tree
of a 360° LiDAR (e.g., Velodyne Puck LITE), the vertical FoV is canopy, cave, tunnel, and post-disaster buildings) and involve both
±15°, leading to a solid angle of 3.25 sr at static. If the LiDAR is in- static structures and dynamic objects (e.g., birds and animals). As
stalled on PULSAR vertically, considering some area that is blocked demonstrated in the experiments, PULSAR can operate safely in
by the UAV structure, an FoV with a solid angle of 10.66 sr (84.8% these real-world environments and acquire environmental data
of omnidirection) can be achieved via the self-rotation, which is 3.28 rapidly, both day and night, giving timely and reliable feedback
times higher than the original FoV. The results quantified the exten- for decision-making.
sion of FoV through self-rotation. In the current implementation of
PULSAR, the Livox AVIA LiDAR was used because of its long
sensing distance and small weight (498 g). MATERIALS AND METHODS
Working principle of the swashplateless mechanism
Agility The working principle of the swashplateless mechanism is illustrat-
PULSAR achieved a good level of agility, as demonstrated in the ex- ed in Fig. 2 (E and F) and movie S1. Because of the two tilted passive
periments of trajectory tracking and step position response. This hinges, the blade pitch can be changed cyclically by modulating an
enabled PULSAR to perform challenging tasks like avoiding high- acceleration impulse to the motor at each rotor revolution. When
speed dynamic obstacles, which are possible only in quadrotor the impulse is applied, the rotor starts accelerating from its
UAVs (46, 47) to date. For existing self-rotation UAVs (20–33) present position (i.e., angle α) to a new position (i.e., angle
and other underactuated UAVs with fewer than four propellers α + λ0). Because of the blade's inertia, its rotation will lag from
(40, 48–52), the main focuses were on design concept, vehicle con- the rotor during the rotor acceleration by rotating along the
figuration, and flying feasibility, whereas agility and the ability to hinge, leading to a lag angle δ. Then, because of the 45°-tilted and
avoid dynamic obstacles were not considered or demonstrated. asymmetric hinge directions, the lag angle causes the positive blade
Only the dSAW proposed in (28) demonstrated an agile motion (the red one) to twist below the propeller disk, leading to an in-
that involved only unpowered diving. creased pitch angle, whereas the other blade (the blue one) twists
Although it has a good level of agility, PULSAR has a thrust-to- above the disk, leading to a decreased pitch angle (Fig. 2F). The dif-
weight ratio that is obviously lower than those of quadrotors. For the ferential changes of the blade pitch angles then cause unequal
two benchmarked quadrotors, their four propellers can provide thrusts on the two blades, which lastly produce a net momentary
much greater total thrust at the cost of high power consumption, moment perpendicular to the blades' feathering axes (the purple
leading to a thrust-to-weight ratio of up to 2.7 (7.5-inch propellers) arrow in Fig. 2E). Similarly, when the rotor is at an opposite posi-
and 3.5 (8-inch propellers), which are 1.29 and 1.67 times higher tion, a deceleration impulse is applied; the decelerating rotor then

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 12 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

causes a decreased pitch angle on the positive blade and an in- According to the angular momentum theorem, the derivative of
creased pitch angle on the negative blade. The net effect is LI is equal to the exerted moment MI ¼ RIB MB , which implies
another momentary moment that has the same direction as the
B
former one. Consequently, during one rotor revolution, two mo- _ Þ
MB ¼ bωB cðIBO ωB þ IBR ΩB Þ þ ðIBO ω_ B þ IBR Ω ð5Þ
mentary moments with the same orientation can be produced to
actuate the UAV attitude. The summation of the two moments where MB ¼ ½MBx ; MBy ; MBprop þ MBbody �, MBprop , and MBbody are the
remains within the propeller disk, with its orientation angle β deter- moments induced by air drag exerted to the rotor part and body
mined by the rotor angle where the acceleration starts (i.e., angle α) part, respectively; MBx and MBy are the moments in the body x and
and its magnitude determined by the magnitude of the impulse. y axis generated by the swashplateless mechanism, respectively; and
Limited by the rotor inertia, impulse acceleration and decelera- ⌊⋅⌋ takes the elements of a vector to form a skew-symmetric matrix.
tion cannot be achieved in practice. Instead, a smoother sinusoidal According to the previous analysis,
motor speed profile is adopted. To produce a sinusoidal motor
speed, the throttle command ut is designed as MBx ¼ MC cosβ; MBy ¼ MC sinβ ð6Þ

ut ¼ ua þ us sinðφ αÞ ð1Þ where MC is the moment magnitude determined by Eq. 2 and β is

the moment orientation angle determined by Eq. 3.
where ua is the average throttle to maintain an average motor speed, Because PULSAR is designed to rotate freely along its body z
us is the amplitude of the sinusoidal throttle, and ϕ is the current axis, we focus on the angular dynamics in the x and y axes only,
rotor angle measured by the magnetic encoder. obtaining (see detailed derivation in the Supplementary Materials)
In Eq. 1, the average throttle ua and the amplitude of sinusoidal
throttle us determine the propeller thrust and moment, respectively, MBxy ¼ IBO ω_ Bxy bLBgyro cωBxy ð7Þ

Downloaded from https://www.science.org on March 18, 2023

as below:
T T
f T ¼ ka ua ; M C ¼ ks u s ð2Þ where MBxy ¼ ½MBx ; MBy ; 0� , ωBxy ¼ ½ωBx ; ωBy ; 0� , and
T
where fT is the propeller thrust and MC is the magnitude of the LBgyro ¼ ½0; 0; ðI BO;z I BO;y ÞωBz
þ I BR;z σ� .
moment. The linear relations in Eq. 2 hold at the hovering condition The kinematics model contains rotation motion and linear
with roughly constant coefficients ka and ks determined from the motion. Instead of describing the full rotation RI B , we only need
data of the swashplateless mechanism in the Supplementary Mate- to describe the kinematics of the z axis of the body frame B (i.e.,
rials (fig. S2). zIB ) because the rotation along zIB is free, and only the rotation per-
Last, the rotor acceleration position α in Eq. 1 determines the pendicular to zIB will affect the UAV’s linear motion. As proved in
orientation angle of the moment (Fig. 2E), as below: the Supplementary Materials, the rotation kinematics is
β ¼ α þ λ0 δ þ π=2 � α þ λ0 þ π=2 ð3Þ z_ IB ¼ ωIxy � zIB ð8Þ
In practice, the lag angle δ is very small and can be deemed as IB
zero, and the angle λ0 is constant and calibrated in advance. where ωIxy ¼R ωBxy . Last, the linear motion is described as
fT I
Dynamic modeling v_ I ¼ z þ gI ð9Þ
m B
PULSAR consists of two rigid parts: One part consists of the UAV
body and motor stator (referred to as the body part), and the other
consists of the propeller and motor rotor (referred to as the rotor p_ I ¼ vI ð10Þ
part). To model the UAV dynamics, we defined two coordinate I
where m is the total mass of PULSAR, g is the gravity vector, and fT
frames (fig. S11): The first one is an inertial (ground-fixed) coordi-
is the propeller thrust determined by Eq. 2.
nate frame I whose z axis is the same as the direction of the gravity
vector, and the second one is a body-fixed coordinate frame B at-
Trajectory tracking control
tached to the body part with origin at the CoM of the whole UAV
The trajectory tracking control of PULSAR adopts a standard dual-
and aligned with the principal axis of inertia. Let
T loop cascaded controller structure (fig. S12), where the outer-loop
ωB ¼ ½ωBx ; ωBy ; ωBz � be the body’s angular velocity. Then, the position controller is a proportional controller and the inner-loop
angular momentum LI of PULSAR observed in the inertial frame velocity controller is a PID controller. The acceleration feedforward
I is of the trajectory is added, with the acceleration command computed
by the velocity controller to produce the total desired acceleration
LI ¼ RI B ðIBO ωB þ IBR ΩB Þ ð4Þ aId , which is then substituted into Eq. 9 to obtain the desired thrust
where RI B is the rotation matrix and its superscript indicates the fT,d and attitude zIB;d
coordinate transformation from B to I ; ΩB = [0,0, σ]T, with σ
f Id
being the rotor speed with respect to the stator; IBO is the inertia f Id ¼ m ðaId gI Þ; f T;d ¼ kf Id k; zIB;d ¼ ð11Þ
fT;d
matrix of the whole UAV including both rotor and body parts;
and IBR is the inertia matrix of the rotor part only, where the refer- Note that we only have to specify the body z axis of the UAV
ence points O and R are the CoM of the whole UAV and the rotor without restricting its yaw angle, which is free rotating for
part, respectively (fig. S11). PULSAR. To track the desired attitude zIB;d , an angular velocity of

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 13 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

the body frame B that rotates zIB (the current attitude) toward zIB;d which runs on the flight controller board at multiple frequencies:
should be generated. In PULSAR, the desired angular velocity ωIxy;d 50 Hz for position control, 200 Hz for attitude control, 800 Hz
for angular velocity control, and 910 Hz for the mixer. This part
is designed as
has only one output: the total throttle command, which indicates
ωIxy;d ¼ kap ðzIB � zIB;d Þ ð12Þ the single-actuated characteristic of PULSAR. The second part is
the navigation module, which consists of odometry and trajectory
kap
where is a proportional gain of the attitude controller and zIB is planning and runs on the onboard ARM computer at 50 Hz. The
the body z axis estimated by the navigation module. As proved in data flow from the onboard computer to the flight controller is
the Supplementary Materials, when the desired angular velocity the UAV states estimated by the LiDAR-inertial odometry and the
ωIxy;d computed in Eq. 12 is accurately tracked and the gain kap is pos- trajectory commands generated by three means: The preset trajec-
tory library contains trajectories planned offline, the trajectory
itive, zIB converges to the desired direction zIB;d , achieving the
planner plans a smooth and obstacle-free trajectory in real time ac-
control objective. cording to the actual perception of the environment, and the
To track the desired angular velocity ωIxy;d computed in Eq. 12, dynamic obstacle detector and planner detects dynamic obstacles
we transform it into the body frame B where the angular velocity and generates an evasive trajectory. Different experiments will
dynamics is modeled by choose the trajectory command according to the task requirement.
T T The communication framework of software modules was ROS
ωBxy;d ¼ ðRI B Þ ωIxy;d ¼ kap ðRI B Þ bzIB czIB;d ¼ kap bzBB czBB;d ð13Þ
Noetic running in Ubuntu 20.04. All the software modules were im-
which naturally leads ωBxy;d to have components only in x and y plemented in C++. The flight control firmware of PULSAR was de-
T veloped on the basis of PX4 V1.11.2. The communication between
because zBB
¼ ½0; 0; 1� . Then, a PI controller is used to track the

Downloaded from https://www.science.org on March 18, 2023

the onboard computer and the flight controller was based on
desired angular velocity, leading to a control law MAVROS (55).
ðt
~ Bxy þ Kri ω
MBxy;d ¼ Krp ω ~ Bxy dτ ð14Þ LiDAR-inertial odometry
0
The full-state estimation of PULSAR was realized by FAST-LIO2
T
where MBxy;d ¼ ½MBx;d ; MBy;d ; 0� is the desired moment, (56), an efficient and robust LiDAR-inertial odometry framework.
FAST-LIO2 estimates the UAV state (i.e., position, velocity, and at-
~ Bxy
ω ¼ ωBxy;d ωBxy
is the angular velocity error, both are vectors in
titude including the self-rotation) and updates the local map in the
the body frame B, and Krp and Kri are two gain matrices that are inertial (ground-fixed) frame (see the description in the “Dynamic
tuned appropriately. Because both ωBxy;d and ωBxy have a zero z com- modeling” section) whose heading is chosen as the initial UAV
ponent, the computed moment MBxy;d will naturally have a zero z heading. We set the LiDAR scan rate to 50 Hz, which implies a
50-Hz state estimation and map update. Such a high-frequency
component too. This moment, if executed successfully, will lead
map update enables FAST-LIO2 to tightly track even very aggressive
the angular velocity ωBxy , which is subject to the first-order dynamics motions, such as the self-rotation of PULSAR. The estimated state
in Eq. 7, to converge to the desired value ωBxy;d . includes the UAV position, velocity, and attitude, which are fed to
an extended Kalman filter (EKF) implemented on the flight control-
Mixer for driving swashplateless mechanism ler. The EKF will further fuse the states with the onboard inertial
The computed desired thrust fT,d (Eq. 11) and desired moment measurement unit (IMU) to refine the UAV state at a higher fre-
MBxy;d (Eq. 14) are used to generate the total motor throttle quency for the use of the controllers that are implemented on the
command ut in the form of Eq. 1. Specifically, replacing fT with same flight controller. Note that the state estimation of PULSAR’s
fT,d and MC with kMBxy;d k in Eq. 2, the average throttle and the am- EKF is dependent on both the IMU and additional sensor (i.e.,
LiDAR or external motion capture system). Without the additional
plitude of the sinusoidal throttle are
sensor, the estimation does not work properly. The details of the
fT;d kMBxy;d k EKF are described in the Supplementary Materials.
ua ¼ ; us ¼ ð15Þ
ka ks
Incremental k-dimensional forest
and the moment orientation angle is Besides odometry, another fundamental requirement for autono-
mous flight and dynamic obstacle avoidance is a map that represents
β ¼ atan2ðMBy;d ; MBx;d Þ ð16Þ both static and dynamic obstacles in the flying environments. We
which then determines the sinusoidal phase α = β − λ0 − π/2 ac- developed a map, called incremental k-dimensional forest (ikd-
cording to Eq. 3. Last, the computed us, ua, and α along with the forest, see Fig. 3B), which is a collection of incremental k-dimen-
current rotor angle ϕ measured by the magnetic encoder lead to sional trees (ikd-trees) (57), each contained in a voxel (cubic side
the total throttle ut, which is sent to the ESC for execution. length is 1 m). An ikd-tree first downsamples the points in the cor-
responding voxel by retaining only the centermost point according
Software framework overview to a prescribed resolution (0.1 m). Then, the retained points are or-
As shown in Fig. 3A, the software framework is divided into two ganized into a k-dimensional tree structure for efficient nearest
parts running on the respective hardware. The first part is the neighbor search. An ikd-tree has the advantages of incremental
flight control, including the controllers, estimator, and mixer, updates (insert and delete) and dynamic rebalancing, which are
all inherited by the ikd-forest. Furthermore, when compared with

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 14 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

an ikd-tree that builds all points into a large tree, the ikd-forest Figs. S1 to S13
Tables S1 to S5
achieves higher efficiency because the tree size of each ikd-tree in
References (61–64)
each voxel is reduced substantially.
On each node of an ikd-tree, it saves the point coordinates. To Other Supplementary Material for this
distinguish points on dynamic objects from those on static ones, manuscript includes the following:
two extra temporal characteristics are also saved on the node: the Movies S1 to S10
hitting point counter and the last hitting point timestamp. The
hitting points counter records the number of points that are too
close (i.e., within the resolution 0.1 m) to the point on the node REFERENCES AND NOTES
(i.e., the centermost point), although these non-centermost points 1. M. Hasanzade, O. Shadeed, E. Koyuncu, Deep reinforcement learning based aggressive
collision avoidance with limited fov for unmanned aerial vehicles, in AIAA SCITECH 2022
have been removed by the downsampling. The last hitting point Forum (AIAA, 2022), p. 2043.
timestamp denotes the timestamp of the last point hitting the node. 2. Y. Zhao, L. Yan, Y. Chen, H. Xie, B. Xu, Faep: Fast autonomous exploration planner for uav
equipped with limited fov sensor. arXiv:2202.12507 [cs.RO] (25 February 2022).
Trajectory generation and autonomous flight 3. C. M. Eaton, E. K. Chong, A. A. Maciejewski, Robust UAV path planning using pomdp with
To generate a safe trajectory in an environment with obstacles, we limited fov sensor, in 2017 IEEE Conference on Control Technology and Applications (CCTA)
(IEEE, 2017), pp. 1530–1535.
deployed a path planning module on the onboard computer,
4. P. Theodorakopoulos, S. Lacroix, Uav target tracking using an adversarial iterative predic-
running at 50 Hz by receiving the point clouds measured by the tion, in 2009 IEEE International Conference on Robotics and Automation (IEEE, 2009),
onboard LiDAR. The path planning module was implemented on pp. 2866–2871.
the basis of a time-accumulated local map and a kinodynamic A* 5. C. Yu, X. Chen, Leader-follower formation for UAVS with fovs constraint, in 2021 22nd IEEE
search algorithm (58). To achieve a faster nearest neighbor search International Conference on Industrial Technology (ICIT) (IEEE, 2021), vol. 1, pp. 1119–1124.

Downloaded from https://www.science.org on March 18, 2023

for obstacle avoidance in path planning, the ikd-forest was used 6. M. Miiller, F. Steidle, M. J. Schuster, P. Lutz, M. Maier, S. Stoneman, T. Tomic, W. Stürzl, Robust
visual-inertial state estimation with multiple odometries and efficient mapping on an MAV
to maintain the time-accumulated local map that only considers
with ultra-wide FOV stereo vision, in 2018 IEEE/RSJ International Conference on Intelligent
points that appeared in recent scans (indicated by the last hitting Robots and Systems (IROS) (IEEE, 2018), pp. 3701–3708.
point timestamp). Furthermore, a “receding planning horizon” 7. W. Gao, K. Wang, W. Ding, F. Gao, T. Qin, S. Shen, Autonomous aerial robot using dual-
strategy (59) was used where the planner’s starting position was fisheye cameras. J. Field Robot. 37, 497–514 (2020).
taken from the current trajectory that is 20 ms after the current 8. A. Gurtner, D. G. Greer, R. Glassock, L. Mejias, R. A. Walker, W. W. Boles, Investigation of fish-
time. The “receding planning horizon” strategy removes the depen- eye lenses for small-uav aerial photography. IEEE Trans. Geosci. Remote Sens. 47,
709–721 (2009).
dence of the planner on the odometry so that they can run in par-
9. M. Tarhan, E. Altuğ, Ekf based attitude estimation and stabilization of a quadrotor uav
allel. The planned trajectory is then transmitted to the flight using vanishing points in catadioptric images. J. Intell. Robot. Syst. 62, 587–607 (2011).
controller for tracking. Once a collision is detected on the current 10. V. M. Respall, D. Devitt, R. Fedorenko, Unmanned aerial vehicle path planning for explo-
trajectory or the UAV is far from the trajectory under tracking, a ration mapping, in 2020 International Conference Nonlinearity, Information and Robotics
replan module is triggered and a new trajectory will be replanned (NIR) (IEEE, 2020), pp. 1–6.

to ensure flight safety in a complex environment. When the 11. J. Zhu, J. Zhu, X. Wan, C. Xu, Downside hemisphere object detection and localization of mav
by fisheye camera, in 2018 15th International Conference on Control, Automation, Robotics
UAV’s current position is within 1 m of the target waypoint, the and Vision (ICARCV) (IEEE, 2018), pp. 532–537.
waypoint is deemed as arrived, and the next waypoint will be 12. G. Kulathunga, R. Fedorenko, A. Klimchik, Regions of interest segmentation from LiDAR
used as the new target point triggering the above planning process. point cloud for multirotor aerial vehicles, in 2020 International Conference on Unmanned
Aircraft Systems (ICUAS) (IEEE, 2020), pp. 1213–1220.
Dynamic obstacle avoidance 13. A. Harmat, M. Trentini, I. Sharf, Multi-camera tracking and mapping for unmanned aerial
vehicles in unstructured environments. J. Intell. Robot. Syst. 78, 291–317 (2015).
A detector was designed to detect dynamic obstacles approaching
14. D. Wierzbicki, Multi-camera imaging system for UAV photogrammetry. Sensors 18,
the UAV. Because the dynamic obstacles are present only for a 2433 (2018).
short period and appear at different positions in successive scans, 15. P. Gohl, D. Honegger, S. Omari, M. Achtelik, M. Pollefeys, R. Siegwart, Omnidirectional visual
points collected in recent scans (indicated by the last hitting point obstacle detection using embedded FPGA, in 2015 IEEE/RSJ International Conference on
timestamp) and that have small hitting point counters are consid- Intelligent Robots and Systems (IROS) (IEEE, 2015), pp. 3938–3943.
ered dynamic obstacles, whereas the rest of the points are consid- 16. G. Zhou, L. Fang, K. Tang, H. Zhang, K. Wang, K. Yang, Guidance: A visual sensing platform
for robotic applications, in Proceedings of the IEEE Conference on Computer Vision and
ered static obstacles. In experiments, the threshold of the
Pattern Recognition Workshops (2015), (IEEE, 2015), pp. 9–14.
timestamp and the counter was well tuned to achieve a reliable de- 17. S. Zhao, H. Zhang, P. Wang, L. Nogueira, S. Scherer, Super odometry: IMU-centric LiDAR-
tection performance. Once a dynamic obstacle is detected, a target visual-inertial estimator for challenging environments, in 2021 IEEE/RSJ International Con-
point is generated such that its distance to the UAV is shorter than ference on Intelligent Robots and Systems (IROS) (IEEE, 2021), pp. 8729–8736.
the nearest static points in the map and that it lies in a direction 18. B. MacAllister, J. Butzke, A. Kushleyev, H. Pandey, M. Likhachev, Path planning for non-
orthogonal to the object’s incoming direction. The first condition circular micro aerial vehicles in constrained environments, in 2013 IEEE International Con-
ference on Robotics and Automation (IEEE, 2013), pp. 3933–3940.
ensures that the space between the UAV and the target point does
19. M. Karimi, M. Oelsch, O. Stengel, E. Babaians, E. Steinbach, Lola-slam: Low-latency lidar
not have any static obstacles, and the second condition facilitates the slam using continuous scan slicing. IEEE Robot. Autom. Lett. 6, 2248–2255 (2021).
UAV to evade the dynamic obstacles. Last, the generated target 20. S. Jameson, K. Fregene, M. Chang, N. Allen, H. Youngren, J. Scroggins, Lockheed Martin’s
point is sent to the flight controller for onboard control. SAMARAI nano air vehicle: Challenges, research, and realization, in 50th AIAA Aerospace
Sciences Meeting Including the New Horizons Forum and Aerospace Exposition (AIAA,
2012), p. 584.
21. S. Bai, P. Chirarattananon, Design and take-off flight of a samara-inspired revolving-wing
Supplementary Materials robot, in 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
This PDF file includes:
(IEEE, 2019), pp. 6070–6076.
Supplementary Text

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 15 of 16

SCIENCE ROBOTICS | RESEARCH ARTICLE

22. S. Bai, Q. He, P. Chirarattananon, A bioinspired revolving-wing drone with passive attitude 46. D. Falanga, K. Kleber, D. Scaramuzza, Dynamic obstacle avoidance for quadrotors with
stability and efficient hovering flight. Sci. Robot. 7, eabg5913 (2022). event cameras. Sci. Robot. 5, eaaz9712 (2020).
23. A. Safaee, S. Z. Moussavi, M. B. Menhaj, Design and construction of monocopter and its 47. B. He, H. Li, S. Wu, D. Wang, Z. Zhang, Q. Dong, C. Xu, F. Gao, Fast-dynamic-vision: Detection
nonlinear control using photo diode array. U. Porto J. Eng. 4, 34–41 (2018). and tracking dynamic objects with event and depth sensing, in 2021 IEEE/RSJ International
24. E. R. Ulrich, D. J. Pines, J. S. Humbert, From falling to flying: The path to powered flight of a Conference on Intelligent Robots and Systems (IROS) (IEEE, 2021), pp. 3071–3078.
robotic samara nano air vehicle. Bioinspir. Biomim. 5, 045009 (2010). 48. O. Carholt, E. Fresk, G. Andrikopoulos, G. Nikolakopoulos, Design, modelling and control of
25. K. Fregene, S. Jameson, D. Sharp, H. Youngren, D. Stuart, Development and flight validation a single rotor UAV, in 2016 24th Mediterranean Conference on Control and Automation (MED)
of an autonomous mono-wing UAS, in American Helicopter Society Forum (Vertical Flight (IEEE, 2016), pp. 840–845.
Society, 2010). 49. J. Buzzatto, M. Liarokapis, An agile, coaxial, omnidirectional rotor module: On the devel-
26. J. T. Isaacs, C. Magee, A. Subbaraman, F. Quitin, K. Fregene, A. Teel, U. Madhow, J. Hespanha, opment of hybrid, all terrain robotic rotorcrafts, in 2020 IEEE International Symposium on
GPS-optimal micro air vehicle navigation in degraded environments, in 2014 American Safety, Security, and Rescue Robotics (SSRR) (IEEE, 2020), pp. 162–168.
Control Conference (IEEE, 2014), pp. 1864–1871. 50. Q. Zhang, Z. Liu, J. Zhao, S. Zhang, Modeling and attitude control of bi-copter, in 2016 IEEE
27. D. Sharp, C. Stoneking, K. Fregene, Micro air vehicle based navigation aiding in degraded International Conference on Aircraft Utility Systems (AUS) (IEEE, 2016), pp. 172–176.
environments, in 2016 IEEE/ION Position, Location and Navigation Symposium (PLANS) (IEEE, 51. B. Li, L. Ma, D. Huang, Y. Sun, A flexibly assembled and maneuverable reconfigurable
2016), pp. 305–312. modular multirotor aerial vehicle. IEEE ASME Trans. Mechatron. 27, 1704–1714 (2022).
28. S. K. H. Win, L. S. T. Win, D. Sufiyan, G. S. Soh, S. Foong, An Agile samara-inspired single- 52. W. Hao, B. Xian, T. Xie, Fault-tolerant position tracking control design for a tilt tri-rotor
actuator aerial robot capable of autorotation and diving. IEEE Trans. Robot. 38, unmanned aerial vehicle. IEEE Trans. Ind. Electron. 69, 604–612 (2022).
1033–1046 (2022). 53. J. J. Paulos, M. Yim, Scalability of cyclic control without blade pitch actuators, in 2018 AIAA
29. M. Piccoli, M. Yim, Passive stability of a single actuator micro aerial vehicle, in 2014 IEEE Atmospheric Flight Mechanics Conference (AIAA, 2018), pp. 1–18.
International Conference on Robotics and Automation (ICRA) (IEEE, 2014), pp. 5510–5515. 54. T. Stoffregen, G. Gallego, T. Drummond, L. Kleeman, D. Scaramuzza, Event-based motion
30. M. Piccoli, M. Yim, Piccolissimo: The smallest micro aerial vehicle, in 2017 IEEE International segmentation by motion compensation, in Proceedings of the IEEE/CVF International Con-
Conference on Robotics and Automation (ICRA) (IEEE, 2017), pp. 3328–3333. ference on Computer Vision (IEEE, 2019), pp. 7244–7253.
31. W. Zhang, M. W. Mueller, R. D’Andrea, A controllable flying vehicle with a single moving 55. Mavros,https://github.com/mavlink/mavros.
part, in 2016 IEEE International Conference on Robotics and Automation (ICRA) (IEEE, 2016), 56. W. Xu, Y. Cai, D. He, J. Lin, F. Zhang, FAST-LIO2: Fast direct LiDAR-inertial odometry. IEEE

Downloaded from https://www.science.org on March 18, 2023

pp. 3275–3281. Trans. Robot. 38, 2053–2073 (2022).
32. L. S. T. Win, S. K. H. Win, D. Sufiyan, G. S. Soh, S. Foong, Achieving efficient controlled flight 57. Y. Cai, W. Xu, F. Zhang, ikd-Tree: An incremental K-D tree for robotic applications.
with a single actuator, in 2020 IEEE/ASME International Conference on Advanced Intelligent arXiv:2102.10808 [cs.RO] (22 February 2021).
Mechatronics (AIM) (IEEE, 2020), pp. 1625–1631. 58. F. Kong, W. Xu, Y. Cai, F. Zhang, Avoiding dynamic small obstacles with onboard sensing
33. S. K. H. Win, L. S. T. Win, D. Sufiyan, S. Foong, Design and control of the first foldable single- and computation on aerial robots. IEEE Robot. Autom. Lett. 6, 7869–7876 (2021).
actuator rotary wing micro aerial vehicle. Bioinspir. Biomim. 16, 066019 (2021). 59. A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, R. Siegwart, Receding horizon path planning
34. S. Sun, G. Cioffi, C. De Visser, D. Scaramuzza, Autonomous quadrotor flight despite rotor for 3d exploration and surface inspection. Auton. Robots 42, 291–306 (2018).
failure with onboard vision sensors: Frames vs. events. IEEE Robot. Autom. Lett. 6, 60. N. Chen, Data of PULSAR, Zenodo (2023); https://doi.org/10.5281/zenodo.7656487.
580–587 (2021). 61. APC propeller database,https://www.apcprop.com/technical-information/performance-
35. J. Paulos, M. Yim, An underactuated propeller for attitude control in micro air vehicles, in data/.
2013 IEEE/RSJ International Conference on Intelligent Robots and Systems (IEEE, 2013), 62. B. W. McCormick, Aerodynamics, Aeronautics, and Flight Mechanics (John Wiley &
pp. 1374–1379. Sons, 1994).
36. J. Paulos, M. Yim, Flight performance of a swashplateless micro air vehicle, in 2015 IEEE 63. H. Glauert, Aerodynamic Theory (Springer, 1935), pp. 169–360.
International Conference on Robotics and Automation (ICRA) (IEEE, 2015), pp. 5284–5289.
64. R. W. Deters, G. K. Ananda Krishnan, M. S. Selig, Reynolds number effects on the perfor-
37. H. Huang, G. M. Hoffmann, S. L. Waslander, C. J. Tomlin, Aerodynamics and control of au- mance of small-scale propellers, in 32nd AIAA Applied Aerodynamics Conference (AIAA,
tonomous quadrotor helicopters in aggressive maneuvering, in 2009 IEEE International 2014), p. 2151.
Conference on Robotics and Automation (IEEE, 2009), pp. 3277–3282.
38. J. Paulos, B. Caraher, M. Yim, Emulating a fully actuated aerial vehicle using two actuators, Acknowledgments: We would like to thank Y. Ren and G. Lu for help with the mechanical
in 2018 IEEE International Conference on Robotics and Automation (ICRA) (IEEE, 2018), design and helpful discussions. Funding: This work was supported by the Hong Kong Research
pp. 7011–7016. Grants Council (RGC) General Research Fund (GRF) (no. 17206920), the Hong Kong Research
39. S. Gudmundsson, General Aviation Aircraft Design: Applied Methods and Procedures (But- Grants Council (RGC) Early Career Scheme (ECS) (no. 27202219), and a DJI research donation.
terworth-Heinemann, ed. 2, 2021). Author contributions: N.C. and W.X. proposed the initial idea of the research. N.C. designed
40. Y. Qin, W. Xu, A. Lee, F. Zhang, Gemini: A compact yet efficient Bi-copter UAV for indoor and manufactured the UAV prototype, formulated the dynamic model with F.Z., and
applications. IEEE Robot. Autom. Lett. 5, 3213–3220 (2020). implemented all software modules for the UAV with the help of F.K., W.X., and Y.C. The
41. D. Barcelos, A. Kolaei, G. Bramesfeld, Aerodynamic interactions of quadrotor configura- experiments were designed by N.C. and performed by N.C., F.K., W.X., Y.C., D.H., and Y.Q. N.C.
tions. J. Aircr. 57, 1074–1090 (2020). and H.L. contributed to the data acquisition on a test stand. N.C. finished all data analyses. N.C.
42. D. Shukla, N. Komerath, Multirotor drone aerodynamic interaction investigation. Drones 2, and F.Z. wrote the manuscript with the advice of all authors. F.Z. provided funding and
43 (2018). supervised the research. Competing interests: The authors declare that they have no
43. W. Zhou, Z. Ning, H. Li, H. Hu, An experimental investigation on rotor-to-rotor interactions competing interests. Data and materials availability: All data needed to evaluate the
of small UAV propellers, in 35th AIAA Applied Aerodynamics Conference (AIAA, conclusions in the paper are present in the paper or the Supplementary Materials and in an
2017), p. 3744. online database (60). The design files and source code are also available on GitHub at https://
44. D. Verstraete, R. MacNeill, The effects of blockage on the performance of small propellers, github.com/hku-mars/PULSAR.
in 20th Australasian Fluid Mechanics Conference (Australasian Fluid Mechanics
Society, 2016). Submitted 19 August 2022
Accepted 21 February 2023
45. P. Foehn, A. Romero, D. Scaramuzza, Time-optimal planning for quadrotor waypoint flight.
Published 15 March 2023
Sci. Robot. 6, eabh1221 (2021).
10.1126/scirobotics.ade4538

Chen et al., Sci. Robot. 8, eade4538 (2023) 15 March 2023 16 of 16

 A self-rotating, single-actuated UAV with extended sensor field of view for
autonomous navigation
Nan Chen, Fanze Kong, Wei Xu, Yixi Cai, Haotian Li, Dongjiao He, Youming Qin, and Fu Zhang

Sci. Robot., 8 (76), eade4538.

DOI: 10.1126/scirobotics.ade4538

Downloaded from https://www.science.org on March 18, 2023

View the article online
https://www.science.org/doi/10.1126/scirobotics.ade4538
Permissions
https://www.science.org/help/reprints-and-permissions

Use of this article is subject to the Terms of service

Science Robotics (ISSN ) is published by the American Association for the Advancement of Science. 1200 New York Avenue NW,
Washington, DC 20005. The title Science Robotics is a registered trademark of AAAS.
Copyright © 2023 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim
to original U.S. Government Works