Supplementary Materials for

Safety-assured high-speed navigation for MAVs

Yunfan Ren et al.

Corresponding author: Fu Zhang, [email protected]

Sci. Robot. 10, eado6187 (2025)

DOI: 10.1126/scirobotics.ado6187

The PDF file includes:

Methods
Figs. S1 to S13
Table S1
Legends for movies S1 to S4
References (66–75)

Other Supplementary Material for this manuscript includes the following:

Movies S1 to S4
Supplementary Methods
System breakdown of SUPER
This section provides a detailed breakdown of the SUPER system with respect to weights, power
consumption, and CPU usage (Fig. S1). As illustrated in Fig. S1A, the total hardware weight
of SUPER is 1,512 g, comprising the airframe (510 g), battery (432 g), onboard computer (266
g), LiDAR sensor (255 g), and other avionics components (52 g). The power consumption
of SUPER was measured during a typical flight mission in a park (see Fig. S2A), with fully
onboard sensing, planning, and control on and an average speed of 2.84 m/s. In flight mission,
SUPER is commanded to continuously traverse the waypoints following: p1 → p2 → p3 →
p4 → p1 → p5 → p3 → p4 → p1 → p6 → p3 → p4 as shown in Fig. S2B. With a
3300 mAh battery, the drone can perform autonomous navigation among a given sequence of
waypoints until the battery is out. The total measured flight time is approximately 11 minutes
and 24 seconds with a total flight distance of 1,942.56 m (see Fig. S2). The total average power
consumption is around 432 W, with 87.9% (379.7 W) consumed by the actuators. The onboard
computer is the second-largest power consumer, drawing around 43.3 W. The LiDAR sensor
and other avionics components consume 6.3 W and 3.2 W, respectively.
The CPU usage of the onboard computer was also measured during the flight mission. The
average CPU usage is around 13.9%, with a peak of 19.7%. The state estimation module is
the primary consumer, taking up 11.1% on average. The three main software components, state
estimation, planning, and control, are running in parallel as separate processes and communicate
with each other through the publisher-subscriber mechanism of the Robot Operating System
(ROS) (66). To further enhance computational efficiency, the processes within each module are
parallelized using multiple threads. In the state estimation module, the nearest neighbor search
during scan registration is parallelized across 16 threads. In the planning module, the updates to
the spatio-temporal sliding map and the trajectory planning tasks, which include both frontend
path search and backend trajectory optimization, are handled by two separate threads. Lastly,
in the control module, the model predictive controller (MPC) operates on a single thread.

Benchmark comparison on LiDAR-based navigation methods

To compare the performance of the planning module of SUPER and state-of-the-art LiDAR-
based navigation system Falco (42), Bubble planner (13) and IPC (67), this section provides a
detailed benchmark comparison in simulation. The setup of the simulation test is the same as
in the main results “Evaluation of Safety, Success Rate, and Efficiency” of the manuscript. We
have tested two versions: Two versions of Falco were tested. The first, Falco without Gimbal,
corresponds to the official open-source implementation of Falco, which does not include a gim-
bal structure as used in their real-world experiments. In this setup, when the drone accelerates
or decelerates, the LiDAR field of view (FOV) and corresponding motion primitives are aligned
with the UAV’s body frame, as shown in Fig. S3A. The second version, Falco with Gimbal,
involves a modification of the official Falco implementation by incorporating a gimbal structure
in the simulation. This modification ensures that the LiDAR sensor’s horizontal orientation re-
mains independent of the drone’s attitude, as illustrated in Fig. S3B. This gimbal configuration,
which is also used in the real-world experiments of Falco, facilitates stable operation during
high-speed flights.
The results are shown in Fig. S4, where the result of Bubble is directly drawn from the main
results. As can be seen, the Falco with Gimbal approach achieves an overall success rate of
57.04% and a maximum attainable average speed exceeding 10 m/s, which is in agreement with
the performance presented in the original paper (42). In contrast, the Falco without Gimbal con-
figuration, the default open-source Falco implementation, exhibits a lower overall success rate
of only 40.28% and a maximum average speed below 6 m/s, even in the sparsest environment.
Regarding IPC, it achieves a high success rate of 75%, outperforming both versions of Falco
and the Bubble planner. Although it shows a relatively lower success rate in low traversability
environments due to its optimistic planning strategy, the success rate increases as traversabil-
ity improves. However, in terms of maximum achievable speed, IPC falls behind both Bubble
and SUPER. These results demonstrate that SUPER achieves a higher success rate compared to
Falco, Bubble and IPC, while attaining much higher average flight speeds. The comparison of
computation times is illustrated in Fig. S4C. Falco, with its offline trajectory library and map-
less planning framework, requires less computation time than both Bubble and SUPER. For
IPC, it generates only one or two polyhedra in each replan cycle, leading to lower computation
costs on the frontend. Additionally, its MPC-based planning and control framework takes less
than 1 ms for the backend, resulting in much lower overall computational costs.

Implementation Details of Different Baseline Methods

We provide a concise comparison of the implementation details of different baseline methods
in the “Evaluation of Safety, Success Rate, and Efficiency” section of the main text. Table S1
presents the comparison of these methods across four aspects. Firstly, the methods are cate-
gorized into three types based on their safety strategies: optimistic, safety-aware, and safety-
assured. Bubble (13) is an optimistic method that treats unknown spaces as free. Raptor (20)
is a safety-aware method that actively plans trajectories to enhance the visibility of unknown
spaces but lacks safety guarantees. Faster (23) and our proposed SUPER are both safety-assured
methods with theoretical safety guarantees. Secondly, the mapping approaches differ among the
methods. Raptor employs an Euclidean signed distance field (ESDF)-based map (31), as Faster
uses an occupancy grid map (OGM). In contrast, Bubble and SUPER adopt a point cloud-based
method, with Bubble additionally utilizing a K-dimensional tree map for corridor generation.
Generally, the point cloud-based mapping approach in Bubble and SUPER exhibits much lower
computation time compared to Faster and Raptor. All four baselines adopt a frontend and
backend style in trajectory generation, with different methods employed for each. Bubble,
Faster, and SUPER utilize flight corridor-based frontends, as Raptor employs topological path
search. For the backend, Faster uses a mixed-integer quadratic programming (MIQP)-based
method, and Raptor uses path-guided trajectory optimization. However, both Faster’s MIQP-
based method and Raptor’s path-guided optimization only optimize the spatial properties of the
trajectory and have limited freedom to find feasible trajectories during high-speed flight, which
necessitates temporal optimization. In contrast, both Bubble and SUPER incorporate spatio-
temporal trajectory optimization based on MINCO (29), demonstrating superior performance
in high-speed flights, as verified in the main text.

Evaluation of the Planning Module of Faster and SUPER

We evaluated the planning module of Faster (23) and SUPER in a simulation, using a challeng-
ing indoor corner environment where an obstacle is concealed behind the corner, as depicted in
Fig. S6A-i. Additionally, we conducted an ablation study on SUPER, specifically SUPER with-
out OT, which is a variant of SUPER that disables switching time optimization. In each test,
the drone commenced at ps = [0, 0, 1.2]T , with the goal set to pg = [12.5, 7.5, 1.2]T . All three
methods achieved the goal safely. Both Faster and SUPER without OT switched to a backup
trajectory twice when the drone detected the obstacle behind the corner. However, due to the
computationally demanding MIQP-based trajectory optimization employed by Faster, it failed
to plan a new trajectory transition from the backup trajectory and execute the backup trajectory
till a stop. Consequently, the execution time of the backup trajectory was much longer for Faster
compared to SUPER without OT, resulting in a total time of 6.2 s for Faster, exceeding the 4.9
s recorded for SUPER without OT. Furthermore, the full version of SUPER, incorporating the
proposed backup trajectory switching time optimization approach, did not switch to the backup
trajectory and avoid unnecessary braking maneuvers. SUPER achieved the shortest arrival time
of 4.8 s, showcasing the high-speed capability of the proposed planning module.
Additionally, we conducted a controlled, quantitative analysis of the backup trajectory gen-
eration between Faster and SUPER. As illustrated in Fig. S6C, both planners were provided
with the same exploratory trajectory and backup corridor. The resulting backup trajectories and
switching times ts are depicted by the green curve and orange circle, respectively. Faster em-
ployed a heuristic approach to determine the switch time (23), setting the switching time ts to be
0.45 s later than the drone’s current state. In comparison, SUPER simultaneously optimized the
switch time and backup trajectory, aiming to maximize ts while ensuring the backup trajectory
remained within the backup corridor. This optimization process yielded a switching time of
ts = 0.84 s for SUPER. Consequently, SUPER delayed the switching time by 86% compared
to Faster, allowing more time for the planning module to re-plan new exploratory and backup
trajectories. As a result, the portion of executing the backup trajectory was reduced, as shown
in Fig. S6B, enhancing the overall flight speed and smoothness.
To verify the performance of our planning module when integrated into the complete sys-
tem, we reproduced the aforementioned simulation experiments in a real-world setting, as
depicted in Fig. S6D-F. The drone started at ps = [0, 0, 1.2]T and was assigned a goal at
pg = [12.5, −7.5, 1.2]T , with a maximum speed limit of vmax = 8 m/s, mirroring the simu-
lation setup. As shown in Fig. S6E, at time t1 = 2.1 s, the obstacle was occluded by the wall,
and the generated exploratory trajectory Te1 collided with the obstacle. However, the backup
trajectory Tb1 was confined within the backup corridor, ensuring that in the worst-case scenario,
the drone could safely come to a stop without colliding with any obstacle. At t2 = 2.2 s, the
drone perceived the obstacle for the first time (Fig. S6D, ii) and re-planned a new exploratory
trajectory Te2 to avoid it. Consequently, the initial backup trajectory Tb1 was not executed, and a
new backup trajectory was generated instead. The velocity profile is shown in Fig. S6F, demon-
strating that the drone did not execute any backup trajectory and successfully reached the goal
at a high average speed, highlighting the high-speed and safe properties of SUPER.

Convex Decomposition in Configuration Space

Convex decomposition is a technique used to identify a convex shape around the given seed
(such as a point or a line segment) that can effectively separate a convex region of space from
a set of obstacles. In robotic navigation missions, performing convex decomposition in the
configuration space and plan trajectories in the decomposed spaces (12, 13, 23, 40) is a pop-
ular way to efficiently achieve collision avoidance. Currently, two mainstream methods exist
for generating free polyhedra in the configuration space (C-space). The first method, called
Point Inflation (12, 23, 60), involves inflating points and performing convex decomposition on
the inflated point cloud (Fig. S7A, i). However, the point inflation exponentially increases the
number of point clouds and negatively affects computational efficiency. Additionally, the in-
flated obstacle point clouds often fail to represent the configuration space accurately, leading to
small and sharp polyhedra (see Fig. S7A, i) that may over-constrained the trajectory optimiza-
tion problems. The second method, known as Plane Shrinkage (40, 61), generates the convex
shape directly from the original point cloud and then shrinks it by the robot’s radius to obtain the
free polyhedron in C-space (Fig. S7A, ii). However, this approach focuses on maximizing the
polyhedron before shrinking, which can lead to the resultant polyhedron being small and sharp
after the shrinking process. Additionally, there is a risk that the resulting shrunken polyhedron
may not encompass the seed, thereby limiting its applicability in scenarios where it is crucial
for the seed to reside within the generated polyhedron.
To address those issues, we propose the Configuration-space Iterative Regional Inflation
(CIRI) algorithm. Unlike previous approaches (40, 58, 59, 68, 69) that represent obstacles as
points, we model obstacle points as spheres with a radius equal to the robot’s radius. This
approach allows us to generate free polyhedra in the C-space directly, offering several advan-
tages. Firstly, our algorithm operates directly on the original point clouds without the need
for inflation. This approach reduces the number of points involved, resulting in a substan-
tial improvement in computational efficiency. Secondly, the accurate modeling of obstacles in
the configuration space enables our method to find polyhedra with larger sizes compared with
the shrinkage-based or inflation-based methods. Additionally, we propose a plane adjustment
method that ensures the inclusion of the seed in the generated polyhedra, which is helpful when
generating backup corridors.
The proposed CIRI builds upon the iterative process of the FIRI framework (58) (see Al-
gorithm 1), offering improved computational efficiency and generating larger polyhedra. The
algorithm takes a line segment seed S : {sa , sb } and the obstacle point cloud O as input, and
begins with the initialization of the ellipsoid as a unit ball (line 2) and proceeds iteratively
until the ellipsoid’s volume converges. The iterative process comprises two steps. First, the
ellipsoid is uniformly inflated until it tangentially intersects the obstacle point cloud, generat-
ing a set of hyperplanes that separate a convex region from the obstacles while ensuring the
inclusion of the seed (line 7). Second, a Second-Order Cone Programming (SOCP)-based op-
timization is applied to identify the maximum-volume ellipsoid within the polyhedron formed
by the obstacle-free half-spaces defined by those hyperplanes (line 8). Through iterative re-
finement, the hyperplanes and the ellipsoid are adjusted to progressively expand the volume of
the inscribed ellipsoid, resulting in a larger polyhedron and an increased region of obstacle-free
space.
Ellipsoid-based Convex Decomposition
In this step, our objective is to determine a free polyhedron comprising a collection of sepa-
rating hyperplanes. These hyperplanes ensure that the given seed is devoid of obstacles. Each
hyperplane is denoted by H(A, b), where Ax ≤ b, x ∈ R3 defines a half-space. The polyhedron
P is defined by intersection of K half-spaces: P : {H1 , . . . , Hk }. To represent the ellipsoid,
we use the image of the unit ball, which is known as the Löwner-John ellipsoid (70), defined as
follows:
E(C, d) = {Cx̃ + d| ∥x̃∥ ≤ 1}, (S1)
 Algorithm 1: Configuration-space Iterative Regional Inflation (CIRI)
Notation: Identity matrix: I ∈ R3×3 ; Ellipsoid E centered at d with the shape matrix C;
Converge threshold: ε
Input : Obstacle point cloud: O; Robot’s radius: r;
Seed: S
Output : Polyhedron: P
1 Function CIRI(O, r, S)
2 E.C = I;
3 E.d = (sa + sb )/2;
4 converge = False;
5 while not converge do
6 Clast = E.C;
7 P = EllipsoidCvxDecomp(O, r, S, E) ;
8 E = MaximumInscribeVolumeEllipsoid(P);
9 if [det(E.C) − det(Clast )] /det(Clast ) < ε then
10 converge = True;
11 end
12 end
13 return(P);
14 End Function

where C ∈ R3×3 is a diagonal shape matrix defined by C = diag(rx , ry , rz ) and d ∈ R3 is the

center of the ellipsoid. We also define its uniform expansion as:

Eα (C, d) = {Cx̃ + d| ∥x̃∥ ≤ α}, (S2)

where α ≥ 1 is the expansion factor. We use the same definition to describe a sphere, where
C = diag(r, r, r) and d is the sphere’s center.
The input of the proposed method consists of the obstacle point cloud O ∈ R3×N with N
points, the ellipsoid in the world frame EW , the seed S and the robot’s radius r. The process is
outlined in Algorithm 2 and visualized in Fig. S7C.
At first, the activating obstacle point clouds Oa are initialized with the full obstacle point
clouds O (line 2). Then, we continued to generate hyperplanes to separate Oa until it became
empty. In each iteration, we choose the nearest obstacle sphere, denoted as EW s , to the center
of the ellipsoid E (line 4) and try to find the separating hyperplanes between the EW s and the
ellipsoid. The key to finding these hyperplanes involves locating the planes that intersect the
obstacles and are tangent to uniformly expanded ellipsoid EW α . Building upon the concept
presented in (59), we simplify the problem by transforming it to the ellipsoid frame E and form
it into a single least-distance programming problem, thus avoiding searching over different
expansion factors α. In Eq. S1, we define the ellipsoid in world frame W as the image of
the unit ball in the ellipsoid frame E. Thus, we construct the inverse map and transform the
problem to the ellipsoid frame E as shown in Fig. S7B. The intersecting point pW E
t ’s image pt is
the closest point to the origin of the E frame on EEs . Now, the problem is turned into a minimum
distance optimization problem.
pEt = arg min ∥x∥2 ,
x
(S3)
s.t. x ∈ EEs .
 Algorithm 2: Ellipsoid-based Convex Decomposition
Params : Active obstacle point cloud Oa ;
Input : Obstacle point cloud: O; Robot’s radius: r;
Seed: S; Ellipsoid: EW
Output : Polyhedron: P
1 Function GeneratePolytope(O, r, S, E)
2 Oa = O;
3 while not Oa .empty() do
4 EW W
s = FindCloestObstacle(E , Oa );
5 H = ComputeSeparatingPlanes(EW s );
6 foreach s ∈ S do
7 H = CheckAndAdjustPlane(H, s, EW s );
8 end
9 Oa = Oa \ (H ∩ Oa );
10 P.AddPlane(H);
11 end
12 return(P);
13 End Function

This problem can be solved efficiently by finding the root of a 6-th-degree polynomial (71).
We implement it with C++, and on an Intel i7 CPU, the typical solving time is less than 0.5 us.
After finding the tangent point pEt , we can find the half space HE (AE , bE ) = (pEt /∥pEt ∥, −∥pEt ∥)
and the expansion factor α = ∥pEt ∥. Then, we can transform them back to get the HW in the
world frame:
HW (CAE , bE + [CAE ] · d), (S4)
where C, d is the shape matrix and the center of E respectively.
The process above finds a separating hyperplane between the ellipsoid and the obstacle
spheres. However, it is possible that the seed does not lie in the half-space defined by the
generated hyperplanes. To address this issue, we propose a plane adjustment method to ensure
that the generated polyhedron includes the seed (line 6). As depicted in Fig. S7D, we first
verify if both sa and sb lie within the half-space defined by H. Without loss of generality, if the
seed point sa is located outside H, we adjust H to a new hyperplane Hadj that passes through
point sa , which is tangent to the obstacle sphere, and points towards another seed point sb .
This adjustment ensures that the new hyperplane Hadj contains both sa and sb . We then set the
hyperplane generated in this round as the new H, denoted as Hadj . After the plane adjustment
process (if necessary), all obstacle points outside H are excluded from Oa , and H is added to
the polyhedron P. This process is repeated until all obstacle points are excluded.
Maximum Volume Inscribed Ellipsoid
Given the polyhedron (a set of hyperplanes), this step involves updating the ellipsoid by deter-
mining the maximum volume inscribed ellipsoid (MVIE) within the given polyhedron. Assume
the polyhedron consists of K hyperplanes P = {H1 , . . . , HK } and the ellipsoid is define by
Eq. S1, the optimization problem can be formulated as:
 max det(C),
C,d

s.t. sup(Ai Cx̃) + Ai d ≤ bi , ∀i = 1, . . . , K, (S5)

C ⪰ 0,
where C and d are the shape matrix and the center of the ellipsoid, respectively. Ai and bi are the
parameters of the plane equation form of the hyperplane H(A, b) : Ax ≤ b. The optimization
problem Eq. S5 is a well-studied convex optimization problem in computational geometry. In
this paper, we utilize the formulation proposed in (72) to convert Eq. S5 as a SOCP-based
method and can be efficiently solved following the Conic Augmented Lagrangian method (73).
Comparison on Convex Decomposition
We conducted a comparative analysis of convex decomposition methods in three distinct en-
vironments characterized by different obstacle shapes: Pillars, Forest, and Perlin (74) (see
(Fig. S8)). For each obstacle type, we randomly generated 300 scenarios with varying den-
sities of obstacles. To quantify the obstacle density of a scenario, we used the obstacle filling
rate (OFR), which is calculated as the ratio of the total volume of obstacles to the total volume of
the space. The 900 scenarios over different obstacle type and densities lead to an OFR ranging
from 0.0005 to 0.6. For each scenario, we sample a point cloud map with a spatial resolution
of 0.1 m, along with a random line seed that was ensured to be at least 0.2 m (equal to the
robot’s radius) away from the nearest obstacles to ensure the existence of a feasible polyhedron
in the configuration space. The point cloud map and line seeds are then fed to different convex
decomposition methods.
We compared three approaches that perform convex decomposition in the configuration
space (see Fig. S7A). The first approach, Plane Shrinkage, generates a polyhedron on the orig-
inal point cloud and then shrinks each hyperplane by the robot’s radius. The second approach,
Point Inflation, inflates the input point clouds by adding additional points within cubes centered
at each input point, where the cube length is equal to the robot’s radius. Convex decomposition
is then performed on the inflated point cloud, ensuring that the resulting polyhedron is valid
in the configuration space. The third approach, CIRI (Configuration-space Iterative Regional
Inflation), models each input point as a sphere with a radius equal to the robot’s radius, and
then performs convex decomposition on these point spheres. For the first two approaches, we
implemented baseline methods based on two state-of-the-art convex decomposition techniques:
FIRI (58) and RILS (40), resulting in FIRI-Shrinkage, FIRI-Inflation, RILS-Shrinkage, and
RILS-Inflation, respectively.
We evaluated each method in terms of seed containment rate (the ratio of polyhedra that con-
tains the line seed), computation time, and volume of the resulting polyhedra (see Fig. S8B-D).
Considering the seed containment rate (Fig. S8B), CIRI and the point inflation (either FIRI-
inflation or RILS-inflation) achieved 100% seed containment rate in all tests. In contrast, the
plane shrinkage-based methods (either FIRI-shrinkage or RILS-shrinkage) exhibited a much
lower seed containment rate, as they can only ensure the seed is within the polyhedron before
the shrinking process but not after. Regarding the volume of the generated polyhedra (Fig. S8C),
CIRI consistently produced polyhedra with the largest average volume compared to the other
four baselines, thanks to its precise modeling of the configuration space. In terms of com-
putational time (Fig. S8D), CIRI and the two shrinkage-based baselines, FIRI-shrinkage and
RILS-shrinkage, had lower computation times compared to the two inflation-based baselines.
This is because the point inflation introduces more points in the input point cloud, which sub-
stantially increases the computation time. Compared to the shrinkage-based methods, CIRI has
slightly higher computation time due to the need to solve the tangent plane between a sphere
and an ellipsoid, while the other two methods solve the point-ellipsoid tangent plane, which is
computationally easier.
Since plane shrinkage-based methods, including FIRI-Shrinkage and RILS-Shrinkage, do
not guarantee the containment of line seed, they are not suitable for SUPER. Comparing among
the seed-containment-assured methods, CIRI can generate polyhedra with larger volumes while
requiring lower computation time than point inflation methods (FIRI-inflation and RILS-inflation).
In conclusion, CIRI effectively enhances corridor quality within less computational time, thereby
further improving the overall performance of SUPER.

Theorem for Backup SFC Generation

Theorem 1. Let D be an omni-directional and infinitely-dense depth image captured at the
position pc ∈ R3 . If a set S ⊆ R3 satisfies the following conditions, it must lie in the known-
free space:

(a) S is a convex set;

(b) S contains no point from D;

(c) pc ⊆ S.

Proof. Assume that S is not entirely contained within the known-free space. This implies the
existence of a point p with an occupancy state that is either occupied or unknown. In the first
case, if p is in the occupied space, it must be captured in D, which contradicts condition (b). In
the second case, if p is in the unknown space, its depth must be greater than any point d ∈ D
with the same bearing direction. For the point d lies on the line connecting p and pc (the bearing
direction), the convexity of S implies that d ∈ S. This once again contradicts condition (b).
Hence, in both cases, we arrive at a contradiction to condition (b), leading to the conclusion that
S must be entirely contained within the known-free space.

Trajectory Planning with Limited LiDAR FOV

When considering the LiDAR sensor’s limited field of view (FOV), the backup corridor must
intersect with the FOV to ensure that the intersecting space lies entirely within the known-
free area. One most common type of LiDAR sensor has a 360-degree horizontal FOV and
a limited vertical FOV, as shown in Fig. S11A. Assuming the vertical FOV angle is θ < π
(Fig. S11B), the total FOV forms a non-convex shape. To ensure that the intersected backup
corridor is convex, we must select a convex subset of the FOV and intersect it with the convex
polytope generated by CIRI with this subset. The largest convex subset of the 360-degree
horizontal FOV is the intersection of two half-spaces defined by hyperplanes that are tangent to
the FOV’s upper and lower bounds, as shown in Fig. S11A and B. The FOV has infinitely many
such convex subsets, each uniquely determined by a one degree-of-freedom (DOF) heading
direction, h, as illustrated in Fig. S11C. To ensure that the resultant backup corridor contains the
largest possible exploratory trajectory, we align h with the tangent direction of the exploratory
trajectory (Fig. S11D).
 To investigate the influence of the vertical FOV angle θ on planning performance, we con-
ducted a series of experiments in a simulated environment with vertical FOV angles ranging
from 30 degrees to 180 degrees (omni-directional FOV). The testing environment is similar to
that used in the benchmark comparisons in the main manuscript. We selected two types of
simulation environments with different traversability scores: one with a traversability of 3.1, in-
dicating a dense environment, and another with a traversability of 6.5, representing a relatively
sparse environment. For each traversability setting, we tested 10 maps, running 10 trials per
map without limiting the maximum velocity. For each trial, we recorded the average speed, the
ratio of executing the backup trajectory, and the final outcome (success or failure). The results
are presented in Fig. S12.
Considering the success rate, SUPER achieved a 100% success rate in both environments,
regardless of lower or higher traversability (Fig. S12A, i and Fig. S12B, i), highlighting its
safety-assured properties. In the denser environment, the average speed increased as the ver-
tical FOV became larger (Fig. S12A, ii), which is due to the fact that a larger FOV enables
generation of larger volume backup corridors, reducing the need to execute the backup trajec-
tory (Fig. S12A, iii), resulting in a higher average speed. In the sparser environment, the average
flight speed also showed a similar trend, though the influence of FOV size was less significant
(Fig. S12B, ii). This is because, in environments with fewer obstacles, even a small FOV is suf-
ficient for planning safe, high-speed trajectories, and the ratio of executing the backup trajectory
is less affected by the FOV size (Fig. S12B, iii).

On-manifold MPC with Rotor Drag Compensation

In our prior work (65), we have proposed an on-manifold model predictive control (OMMPC)
framework for robotic systems and demonstrated indoor quadrotor aerobatics. However, the
aerodynamic effects were neglected in the previous work, which affects the tracking accuracy
when the vehicle performs high-speed flights in outdoor environments. To overcome this limi-
tation, we extend our previous quadrotor model by incorporating rotor drag, which is the main
aerodynamic effect causing tracking error (64). We derive the resultant OMMPC and implement
it for all flight experiments in this paper, demonstrating sufficiently high tracking accuracy. The
effectiveness of this improvement is also verified through comparative experiments.
Quadrotor dynamics
The quadrotor coordinate directions X, Y , and Z are defined as forward, right, and down,
respectively. The quadrotor dynamics involving rotor drag (64) are formulated as follows:
ṗ = v,
1
v̇ = g − aT Re3 − RDRT v, (S6)
m
Ṙ = R⌊ω B ⌋,

where m is the mass of the drone, g represents the gravity vector with a fixed length of 9.81 m/s2 ,
and aT denotes the scalar thrust acceleration. The position and velocity in the inertial frame
are denoted by p ∈ R3 and v ∈ R3 , respectively. The attitude of the airframe is repre-
sented by R ∈ SO(3), and ω denotes angular velocity in the body frame. The operation
⌊·⌋ converts a vector into a skew-symmetric matrix. The operation ⌊·⌋ converts a vector into a
skew-symmetric matrix. D denotes the rotor drag coefficient matrix, which is parameterized as
D = diag(dh , dh , dv ).
 As discussed in (64), the above quadrotor dynamic system with rotor drag is differentially
flat. The flat output vector is typically selected as σ = [p, ψ]T , where ψ represents the yaw
angle. This differential flatness property allows us to plan the trajectory of the flat output σ(t)
instead of the entire state trajectory s(t) of the quadrotor UAV.
System linearization
As previously described in our previous work (65), OMMPC takes the quadrotor system into a
compound manifold, and defines the system states and inputs as follows:

M = R3 × R3 × SO(3), dim(M) = 9,
(S7)
x = pI vI R ∈ M, u = aT ω B ∈ R4 ,

 

and the error between measured state x and reference state xd , lying on the manifold M, is
mapped into the local homeomorphic space (an open set in Euclidean space) around each point
xd using local coordinates. Particularly, the state error is defined as follows:
T
δx ≜ xd ⊟ x = δpT δvT δRT ∈ R9 ,


δp ≜ pd ⊟ p = pd − p ∈ R3 ,
(S8)
δv ≜ vd ⊟ v = vd − v ∈ R3 ,
δR ≜ Rd ⊟ R = Log(RT Rd ) ∈ R3 ,

where ⊟ is an encapsulated operator implemented to the mapping ⊟ : M × M → Rn as

mentioned in (65), and the input error is also defined as
T
δu ≜ ud − u = δaT δω T ∈ R4 ,


δaT ≜ aTd − aT ∈ R, (S9)

δω ≜ ωd − ω ∈ R3 ,

and the original quadrotor system can be equivalently expressed by an error system where the
vehicle states are mapped to the local coordinates at each point along the reference trajectory.
Follow the derivation in (65), the resultant linearized error system is given by

δxk+1 = Fxk δxk + Fuk δuk , (S10)

where
 
I3 I3 ∆t 0
T
j k
∆t d T ∆t d T
Fx k =  0 I 3 − Rk DRdk Rk (2I3 − D) aT me3 + DRdk v − Rdk v  ,
 
m m


0 0 Exp −ωkd ∆t
  (S11)
0 0
Fu k = ∆t  −Rdk e3 0 ,

T
0 A ωkd ∆t

where
sin ∥θ∥ ⌊θ⌋2
   
1 − cos ∥θ∥ ⌊θ⌋
A(θ) = I3 + + 1− . (S12)
∥θ∥ ∥θ∥ ∥θ∥ ∥θ∥2
 Therefore, the OMMPC can be given by the current error and reference trajectory, leading
to an efficient QP solution:
N
X −1
∥δxk ∥2Q+∥δuk ∥2R +∥δxN ∥2P ,


min
δuk
k=0
(S13)
s.t. δxk+1 = Fxk δxk + Fuk δuk ,
δuk ∈ δUk , k = 0, · · · , N − 1,

where N is the predictive horizon, and positive-definite diagonal matrices Q, R, P denote the
penalty of the stage state, stage input, and terminal state, respectively. Uk = {δu ∈ R4 |umin −
udk ≤ δu ≤ umax − udk } denotes the constraints for the input error.
Experiment Results
To quantitatively assess the effectiveness of the proposed OMMPC incorporating the rotor-drag
model, we conducted a comparative analysis with the common quadrotor model (65) using the
PX4 SITL software with the Gazebo simulator (75). The simulator provides realistic dynamics
and air resistance simulations. We generated an offline polynomial trajectory on the flat space
denoted as σ(t) with a maximum velocity of 15 m/s. The trajectory consisted of a straight line
starting from p0 = [0, 0, 0]T and ending at pg = [120, 0, 0]T along the x-axis. Initially, we
generated the desired state trajectory using differential flatness under the common quadrotor
model and tracked the trajectory using the quadrotor OMMPC proposed in (65). As shown in
Fig. S13A, the UAV exhibited significant attitude tracking errors, and the position tracking error
was notably influenced by the velocity direction, reaching a maximum of 0.4 m. In contrast, by
employing the quadrotor model with rotor drag described in equation Eq. (S6), the OMMPC
effectively predicted and compensated for the drag, resulting in improved attitude-tracking per-
formance and a reduced position-tracking error of 0.05 m.
Supplementary Figures and Tables

A System breakdown
Other Avionics LiDAR Planning
Others Control
52 g, 3.4% 6.3 W, 1.5% 2.1 %
3.2 W, 0.7% 0.7 %
Airframe LiDAR Actuator Onboard Computer
510 g, 33.7% 379.7 W, 87.9% 43.3 W, 10.0% Free State
255 g, 16.8 %
86.0% Estimation
11.1 %

Weight Power CPU

1512 g ~432 W 12 Cores 16 Theads
Onboard Up to 4.7 Ghz
Computer
266 g, 17.5 %

Battery
432 g, 28.5%

B Hardware components Onboard Computer

Motors
(T-Motor F90)
(Intel NUC12)

Battery
(Dural sky 3300mAh)
Flight control ESC
unit (Nxt PX4) (T-Motor F60A)
3D printing and Solid state LiDAR
carbonfiber rack frame (Livox Mid 360)

Fig. S1. System breakdown of SUPER (A) The weight, power, and CPU usage breakdown of SUPER. (B) The
hardware components of SUPER.

A B

Fig. S2. Scenario and trajectory of the endurance test (A) The testing scenario. (B) The point cloud view and
flying trajectory of the drone during the endurance test.
 A Falco w/o Gimbal B Falco with Gimbal

Ground
Ground

Motion Primitives
FOV

Fig. S3. Illustration of Falco with and without gimbal structure (A) The original version of Falco’s open-
sourced implementation lacked a gimbal structure. In this configuration, when the multirotor accelerates, the
motion primitives are constrained to point downwards, prohibiting the drone from flying forward. (B) The modified
version of Falco with a gimbal structure. The LiDAR sensor maintains a horizontal orientation independent of the
drone’s attitude.

Fig. S4. Evaluation of success rate, and efficiency (A) The success rate of the benchmarked methods at different
obstacle densities (measured as traversibility) and flight speeds. Empty columns indicate that the corresponding
combination of speed and density was not achieved. (B) Flight results of the benchmarked methods in 1080
experiments. (C) Time consumption of the benchmarked methods, with squares representing the mean values and
error bars indicating the standard deviations of the total computation time. Each mean and standard deviation is
computed from 180 tests.
 A Third person view B Side view

LiDAR
FOV

Fig. S5. Visual example of thin wire detection on OGM. (A) Third-person view of the MAV facing a thin wire
(shown in green). (B) Side view of the detection process. The gray box represents the cell in the occupancy grid
map (OGM) that contains the thin wire. The blue lines indicate the LiDAR beams that pass through the cell without
hitting the wire, while the red line indicates the beam that hits the wire. Since most of the beams do not detect the
wire, the cell is incorrectly classified as free space.
Fig. S6. Evaluation of planning module of Faster and SUPER (A) (i) A top-down view of the simulation envi-
ronment, where a red cylinder is hidden behind the corner. (ii-iv) Comparison of the executed trajectories for the
three methods: Faster (23), the proposed SUPER, and SUPER without switching time optimization (referred to as
SUPER without OT). (B) Velocity profile and execution of backup trajectories in the simulation. (C) Switching
time ts of Faster and SUPER given the same exploratory trajectory and backup corridor. (D) (i) Real-world experi-
ment of SUPER, with the hidden obstacle highlighted by a red dashed box. (ii) First-person view from the onboard
camera at t2 . (E) Planned trajectories at time t1 and t2 . Te1 and Tb1 are the exploratory and backup trajectories at
t1 , respectively, and Te2 is the exploratory trajectory at t2 . The backup trajectory at t2 is not displayed.
(F) Velocity profile in the experiment.
Fig. S7. Illustration of the proposed flight corridor generation algorithm (CIRI). (A) Comparison of three
different methods for generating a safe flight corridor in configuration space. (B) Example illustrating the trans-
formation between the ellipsoid frame E and the world frame W. (C) Visual depiction of the process of convex
decomposition in the C-space. (D) Example demonstrating the plane adjustment strategy to ensure that the given
seeds lie within the generated polyhedron.
 A

Pillars OFR 0.01~0.6 Forest OFR: 0.005 ~ 0.3 Perlin OFR 0.2~0.6

CIRI
FIRI-Inflation
RILS-Inflation
FIRI-Shrinkage
RILS-Shrinkage

Fig. S8. Comparison of convex decomposition in configuration space. (A) Example benchmark environments
with obstacle filling rates (OFR) ranging from 0.005 to 0.6. (B) Seed containment rate of generated polyhedra
by different methods across OFR. (C) Boxplot showing the average polyhedron volumes generated by different
methods across OFR, with each plot based on 100 tests. The central bar represents the median. (D) Boxplot of
computation times for the various methods across OFR, with each plot derived from 100 tests. The central bar
indicates the median.
Fig. S9. The plot of the barrier function Lµ . The plot illustrates the function Lµ . As the parameter µ decreases
from µ = 0.5 to µ = 0.01, the barrier function becomes progressively harder, with a sharper shape.

... - 3.0 - 2.0 2.0 3.0 ...

Fig. S10. The plot of the mapping from η to ts . The plot illustrates a sigmoid-like function
that maps η ∈ R to the interval (tc , to ).
Fig. S11. Backup corridor generation of limited FOV (A) The largest convex subset of a non-convex FOV is
formed by the intersection of two half-spaces defined by the hyperplanes tangent to the FOV’s upper boundary
Hu and lower boundary Hl . (B) Side view of the hyperplanes Hu and Hl along with the vertical FOV angle,
θ. (C) Top-down view showing the convex subset and the two hyperplanes, which are uniquely determined by
the heading direction h. (D) The heading direction is aligned with the exploratory trajectory to ensure that the
generated backup corridor contains the initial portion of the exploratory trajectory.
 A(i) A(ii) A(iii) Traversability: 3.1
A(ii) A(iii)

Backup Trajectory Ratio

Average Speed [m/s]
B(ii) B(iii)
Success Rate

Vertical FOV [Deg.] Vertical FOV [Deg.] Vertical FOV [Deg.]

B(i) B(ii) B(iii) Traversability: 6.5

Backup Trajectory Ratio

Average Speed [m/s]
Success Rate

Vertical FOV [Deg.] Vertical FOV [Deg.] Vertical FOV [Deg.]

Fig. S12. Benchmark comparison with different vertical FOV (A) Benchmarking results on simulation maps
with a traversability score of 3.1. (B) Benchmarking results on simulation maps with a traversability score of
6.5. For both (A) and (B), the distributions (ii-iii) illustrate the average flight speed and the ratio of executing the
backup trajectory. Error bars represent the standard deviation, while the black points indicate the mean values.
Each distribution is based on data from 180 tests.
 Common Model Drag Effect Aware Model

Speed [m/s]

Measurement Measurement
Reference Reference
Pitch [deg.]

Measurement Measurement
Reference Reference
Position Error [m]

Error Posi. X Error Posi. X

Error Posi. Y Error Posi. Y
Error Posi. Z Error Posi. Z

time [s] time [s]

Fig. S13. Comparison of tracking performance with different drone models of MPC. The plot of speed, pitch
angle, and position tracking error in each axis.
 Table S1. Implementation details of different methods

Type Mapping Frontend Backend

Bubble (13) Optimistic Point Cloud + Kd-Tree Sphere-shaped Corridor Spatio-temporal Trajectory Optimization (STTP)
Raptor (20) Safety-aware ESDF Topological Path Searching Path Guided Trajectory Optimization (PGO)
Faster (23) Safety-assured OGM Polyhedron-shaped Corridor Mixed-Integer Qudratic Program (MIQP)
SUPER (ours) Safety-assured Point Cloud Polyhedron-shaped Corridor Spatio-temporal Trajectory Optimization (STTP)
Supplementary Movies

Movie 1. An overview of the proposed SUPER system. SUPER demonstrates its ability to safely navigate
through unknown, cluttered environments at high speeds, avoid thin obstacles like power lines, and perform
robustly in various scenarios, including object tracking and autonomous exploration.

Movie S1. Safe High-speed Navigation in Unknown Environments.

 Movie S2. Navigation in Cluttered Environments.

Movie S3. Demonstration of SUPER’s Ability to Avoid Thin Obstacles.

Movie S4. Applications of SUPER in Object Tracking, Autonomous Exploration, and Waypoint Navigation
Missions.
 REFERENCES AND NOTES

1. J. Verbeke, J. D. Schutter, Experimental maneuverability and agility quantification for rotary

unmanned aerial vehicle. Int. J. Micro Air Veh. 10, 3–11 (2018).

2. B. Mishra, D. Garg, P. Narang, V. Mishra, Drone-surveillance for search and rescue in natural
disaster. Comput. Commun. 156, 1–10 (2020).

3. S. M. S. M. Daud, M. Y. P. M. Yusof, C. C. Heo, L. S. Khoo, M. K. C. Singh, M. S. Mahmood,

H. Nawawi, Applications of drone in disaster management: A scoping review. Sci. Justice 62,
30–42 (2022).

4. B. Rabta, C. Wankmüller, G. Reiner, A drone fleet model for last-mile distribution in disaster
relief operations. Int. J. Disaster Risk Reduct. 28, 107–112 (2018).

5. Y. Song, A. Romero, M. Müller, V. Koltun, D. Scaramuzza, Reaching the limit in autonomous

racing: Optimal control versus reinforcement learning. Sci. Robot. 8, eadg1462 (2023).

6. P. Foehn, A. Romero, D. Scaramuzza, Time-optimal planning for quadrotor waypoint flight. Sci.
Robot. 6, eabh1221 (2021).

7. A. Romero, R. Penicka, D. Scaramuzza, Time-optimal online replanning for agile quadrotor

flight. IEEE Robot. Autom. Lett. 7, 7730–7737 (2022).

8. A. Romero, S. Sun, P. Foehn, D. Scaramuzza, Model predictive contouring control for time-
optimal quadrotor flight. IEEE Trans. Robot. 38, 3340–3356 (2022).

9. E. Kaufmann, L. Bauersfeld, A. Loquercio, M. Müller, V. Koltun, D. Scaramuzza, Champion-

level drone racing using deep reinforcement learning. Nature 620, 982–987 (2023).

10. J. Zhang, R. G. Chadha, V. Velivela, S. Singh, P-cal: Pre-computed alternative lanes for
aggressive aerial collision avoidance, paper presented at the 12th International Conference on
Field and Service Robotics (FSR), Tokyo, Japan, 31 August 2019.

11. J. Zhang, R. G. Chadha, V. Velivela, S. Singh, P-cap: Pre-computed alternative paths to enable
 aggressive aerial maneuvers in cluttered environments, in 2018 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS) (IEEE, 2018), pp. 8456–8463.

12. Y. Ren, S. Liang, F. Zhu, G. Lu, F. Zhang, Online whole-body motion planning for quadrotor
using multi-resolution search, in 2023 IEEE International Conference on Robotics and
Automation (ICRA) (IEEE, 2023), pp. 1594–1600.

13. Y. Ren, F. Zhu, W. Liu, Z. Wang, Y. Lin, F. Gao, F. Zhang, Bubble planner: Planning
highspeed smooth quadrotor trajectories using receding corridors, in 2022 IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS) (IEEE, 2022), pp. 6332–
6339.

14. X. Zhou, Z. Wang, H. Ye, C. Xu, F. Gao, Ego-planner: An esdf-free gradient-based local
planner for quadrotors. IEEE Robot. Autom. Lett. 6, 478–485 (2020).

15. H. Ye, X. Zhou, Z. Wang, C. Xu, J. Chu, F. Gao, Tgk-planner: An efficient topology guided
kinodynamic planner for autonomous quadrotors. IEEE Robot. Autom. Lett. 6, 494–501 (2020).

16. A. Loquercio, E. Kaufmann, R. Ranftl, M. Müller, V. Koltun, D. Scaramuzza, Learning high-

speed flight in the wild. Sci. Robot. 6, eabg5810 (2021).

17. H. D. Escobar-Alvarez, N. Johnson, T. Hebble, K. Klingebiel, S. A. Quintero, J. Regenstein, N.

A. Browning, R-advance: Rapid adaptive prediction for vision-based autonomous navigation,
control, and evasion. J. Field Robot. 35, 91–100 (2018).

18. L. Quan, Z. Zhang, X. Zhong, C. Xu, F. Gao, Eva-planner: Environmental adaptive quadrotor
planning, in 2021 IEEE International Conference on Robotics and Automation (ICRA) (IEEE,
2021), pp. 398–404.

19. L. Wang, Y. Guo, Speed adaptive robot trajectory generation based on derivative property of b-
spline curve. IEEE Robot. Autom. Lett. 8, 1905–1911 (2023).

20. B. Zhou, J. Pan, F. Gao, S. Shen, Raptor: Robust and perception-aware trajectory replanning for
quadrotor fast flight. IEEE Trans. Robot. 37, 1992–2009 (2021).
21. H. Oleynikova, Z. Taylor, R. Siegwart, J. Nieto, Safe local exploration for replanning in
cluttered unknown environments for microaerial vehicles. IEEE Robot. Autom. Lett. 3, 1474–
1481 (2018).

22. S. Liu, M. Watterson, S. Tang, V. Kumar, High speed navigation for quadrotors with limited
onboard sensing, in 2016 IEEE International Conference on Robotics and Automation (ICRA)
(IEEE, 2016), pp. 1484–1491.

23. J. Tordesillas, B. T. Lopez, M. Everett, J. P. How, Faster: Fast and safe trajectory planner for
navigation in unknown environments. IEEE Transact. Robot. 38, 922–938 (2021).

24. E. Mueggler, H. Rebecq, G. Gallego, T. Delbruck, D. Scaramuzza, The event-camera dataset

and simulator: Event-based data for pose estimation, visual odometry, and slam. Int. J. Robot.
Res. 36, 142–149 (2017).

25. Livox Technology Company Limited, Livox Mid-360 User Manual (2023);
https://livoxtech.com/mid-360.

26. Wikipedia, Lockheed Martin F-35 Lightning II stealth multirole combat aircraft (2024);
https://en.wikipedia.org/wiki/Lockheed_Martin_F-35_Lightning_II.

27. DJI, DJI Mavic 3 (2024); https://dji.com/mavic-3.

28. J. Tordesillas, B. T. Lopez, J. P. How, Faster: Fast and safe trajectory planner for flights in
unknown environments, in 2019 IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS) (IEEE, 2019), pp. 1934–1940.

29. Z. Wang, X. Zhou, C. Xu, F. Gao, Geometrically constrained trajectory optimization for
multicopters. IEEE Trans. Robot. 38, 3259–3278 (2022).

30. C. Nous, R. Meertens, C. De Wagter, G. De Croon, Performance evaluation in obstacle

avoidance, in 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS) (IEEE, 2016), pp. 3614–3619.
31. L. Han, F. Gao, B. Zhou, S. Shen, Fiesta: Fast incremental Euclidean distance fields for online
motion planning of aerial robots, in 2019 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS) (IEEE, 2019), pp. 4423–4430.

32. B. Tang, Y. Ren, F. Zhu, R. He, S. Liang, F. Kong, F. Zhang, Bubble explorer: Fast uav
exploration in large-scale and cluttered 3D-environments using occlusion-free spheres.
arXiv:2304.00852 [cs.RO] (2023).

33. P. Foehn, E. Kaufmann, A. Romero, R. Penicka, S. Sun, L. Bauersfeld, T. Laengle, G. Cioffi, Y.

Song, A. Loquercio, D. Scaramuzza, Agilicious: Open-source and open-hardware agile
quadrotor for vision-based flight. Sci. Robot. 7, eabl6259 (2022).

34. NVIDIA Corporation, Nvidia Jetson TX2 embedded AI computing device (2024);
https://developer.nvidia.com/embedded/jetson-tx2.

35. X. Zhou, X. Wen, Z. Wang, Y. Gao, H. Li, Q. Wang, T. Yang, H. Lu, Y. Cao, C. Xu, F. Gao,
Swarm of micro flying robots in the wild. Sci. Robot. 7, eabm5954 (2022).

36. Skydio 2 Plus Enterprise, An intelligent flying robot powered by AI (2024);

https://skydio.com/skydio-2-plus-enterprise.

37. 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).

38. Intel, Intel RealSense Depth Camera D435i (2024); https://intelrealsense.com/depth-camera-

d435i.

39. W. Xu, Y. Cai, D. He, J. Lin, F. Zhang, Fast-lio2: Fast direct lidar-inertial odometry. IEEE
Trans. Robot. 38, 2053–2073 (2022).

40. S. Liu, M. Watterson, K. Mohta, K. Sun, S. Bhattacharya, C. J. Taylor, V. Kumar, Planning

dynamically feasible trajectories for quadrotors using safe flight corridors in 3-d complex
environments. IEEE Robot. Autom. Lett. 2, 1688–1695 (2017).
41. F. Gao, W. Wu, W. Gao, S. Shen, Flying on point clouds: Online trajectory generation and
autonomous navigation for quadrotors in cluttered environments, J. Field Robot. 36, 710–733
(2019).

42. J. Zhang, C. Hu, R. G. Chadha, S. Singh, Falco: Fast likelihood-based collision avoidance with
extension to human-guided navigation. J. Field Robot. 37, 1300–1313 (2020).

43. D. Cheng, F. C. Ojeda, A. Prabhu, X. Liu, A. Zhu, P. C. Green, R. Ehsani, P. Chaudhari, V.

Kumar, Treescope: An agricultural robotics dataset for lidar-based mapping of trees in forests
and orchards, in 2024 IEEE International Conference on Robotics and Automation (ICRA)
(IEEE, 2023), pp. 14860–14866.

44. P. De Petris, H. Nguyen, M. Dharmadhikari, M. Kulkarni, N. Khedekar, F. Mascarich, K.

Alexis, Rmf-owl: A collision-tolerant flying robot for autonomous subterranean exploration, in
2022 International Conference on Unmanned Aircraft Systems (ICUAS) (IEEE, 2022), pp. 536–
543.

45. Y. Cai, F. Kong, Y. Ren, F. Zhu, J. Lin, F. Zhang, Occupancy grid mapping without raycasting
for high-resolution LiDAR sensors. IEEE Trans. Robot. 40, 172–192 (2023).

46. Ouster, Ouster OS1-128 sensor (2024); https://ouster.com/products/scanning-lidar/os1-sensor.

47. DJI, DJI Matrice 300 RTK (2024); https://enterprise.dji.com/zh-tw/matrice-300.

48. LiDAR USA, LiDAR USA surveyor 32 (2024); https://lidarusa.com/.

49. Intel, Product Family D400 Series Datasheet (2024);

https://intelrealsense.com/download/21345/?tmstv=1697035582.

50. A. Hornung, K. M. Wurm, M. Bennewitz, C. Stachniss, W. Burgard, Octomap: An efficient

probabilistic 3d mapping framework based on octrees. Autonom. Robot. 34, 189–206 (2013).

51. Y. Ren, Y. Cai, F. Zhu, S. Liang, F. Zhang, Rog-map: An efficient robocentric occupancy grid
map for large-scene and high-resolution lidar-based motion planning. arXiv:2302.14819 [cs.RO]
 (2023).

52. F. Kong, W. Xu, Y. Cai, F. Zhang, Avoiding dynamic small obstacles with onboard sensing and
computation on aerial robots. IEEE Robot. Autom. Lett. 6, 7869–7876 (2021).

53. H. Wu, Y. Li, W. Xu, F. Kong, F. Zhang, Moving event detection from LiDAR point streams.
Nat. Commun. 15, 345 (2024).

54. T-MOTOR, T-MOTOR F90 series racing drone motor specifications (2024);
https://store.tmotor.com/goods-1064-F90.html.

55. Dronecode Foundation, PX4: Open source autopilot for drone developers (2024); https://px4.io.

56. J. Chen, K. Su, S. Shen, Real-time safe trajectory generation for quadrotor flight in cluttered
environments, in 2015 IEEE International Conference on Robotics and Biomimetics (ROBIO)
(IEEE, 2015), pp. 1678–1685.

57. P. Hart, N. Nilsson, B. Raphael, A formal basis for the heuristic determination of minimum cost
paths. IEEE Trans. Syst. Sci. Cybern. 4, 100–107 (1968).

58. Q. Wang, Z. Wang, C. Xu, F. Gao, Fast iterative region inflation for computing large 2-d/3-d
convex regions of obstacle-free space. arXiv:2403.02977 [cs.RO] (2024).

59. R. Deits, R. Tedrake, Computing large convex regions of obstacle-free space through
semidefinite programming, in Algorithmic Foundations of Robotics XI: Selected Contributions of
the Eleventh International Workshop on the Algorithmic Foundations of Robotics (Springer,
2015), pp. 109–124.

60. L. Yin, F. Zhu, Y. Ren, F. Kong, F. Zhang, Decentralized swarm trajectory generation for lidar-
based aerial tracking in cluttered environments, in 2023 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS) (IEEE, 2023), pp. 9285–9292.

61. J. Ji, N. Pan, C. Xu, F. Gao, Elastic tracker: A spatio-temporal trajectory planner for flexible
aerial tracking, in 2022 International Conference on Robotics and Automation (ICRA) (IEEE,
 2022), pp. 47–53.

62. S. Liu, N. Atanasov, K. Mohta, V. Kumar, Search-based motion planning for quadrotors using
linear quadratic minimum time control, in 2017 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS) (IEEE, 2017), pp. 2872–2879.

63. ZJU FAST Lab, LBFGS-Lite: A header-only L-BFGS unconstrained optimizer (Github, 2024);
https://github.com/ZJU-FAST-Lab/LBFGS-Lite.

64. M. Faessler, A. Franchi, D. Scaramuzza, Differential flatness of quadrotor dynamics subject to

rotor drag for accurate tracking of high-speed trajectories. IEEE Robot. Autom. Lett. 3, 620–626
(2017).

65. G. Lu, W. Xu, F. Zhang, On-manifold model predictive control for trajectory tracking on
robotic systems. IEEE Trans. Ind. Electron. 70, 9192–9202 (2022).

66. A. Koubaa, Ed., Robot Operating System (ROS): The Complete Reference (Volume 1), vol. 625
of Studies in Computational Intelligence (Springer, 2017).

67. W. Liu, Y. Ren, F. Zhang, Integrated planning and control for quadrotor navigation in presence
of suddenly appearing objects and disturbances. IEEE Robot. Autom. Lett. 9, 899–906 (2023).

68. C. Toumieh, A. Lambert, Voxel-grid based convex decomposition of 3D space for safe corridor
generation. J. Intell. Robot. Syst. 105, 87 (2022).

69. X. Zhong, Y. Wu, D. Wang, Q. Wang, C. Xu, F. Gao, Generating large convex polytopes
directly on point clouds. arXiv:2010.08744 [cs.RO] (2020).

70. C. D. Toth, J. O’Rourke, J. E. Goodman, Handbook of Discrete and Computational Geometry

(CRC Press, 2017).

71. D. Eberly, Distance from a point to an ellipse, an ellipsoid, or a hyperellipsoid (Geometric

Tools, 2013); https://geometrictools.com/Documentation/DistancePointEllipseEllipsoid.pdf.

72. Z. Wang, “A geometrical approach to multicopter motion planning,” thesis, Zhejiang

 University, Hangzhou, China (2022).

73. Y. Cui, D. Sun, K.-C. Toh, On the r-superlinear convergence of the KKT residuals generated by
the augmented lagrangian method for convex composite conic programming. Math. Program.
178, 381–415 (2019).

74. W. Wu, mockamap, https://github.com/HKUST-Aerial-Robotics/mockamap.

75. Dronecode Foundation, PX4 software in the loop simulation with Gazebo (2024);
https://docs.px4.io/v1.12/en/simulation/gazebo.html.