In recent years, unmanned aerial vehicles (UAVs) have attracted more and more attention in both academia and industry. Among them, quadrotors are most widely used due to their simple structure, high mobility, and vertical takeoff and landing (VTOL) capabilities. Moreover, progresses in autonomous navigation enable quadrotors to fly safely and aggressively in unknown cluttered environments with full autonomy[25, 26, 24], greatly expanding their application area.
However, quadrotor inherently suffers from bad power utilization (PU) because most of the energy is wasted on counteracting the body weight. This defect limits quadrotors’ use in long-distance missions such as search and rescue, delivery, and active exploration. In such missions, the mobility of quadrotors is necessary for traversing extreme terrain, while the relatively short endurance can hardly support quadrotors to complete the entire mission. In contrast, although unmanned ground vehicles (UGVs) cannot cross rugged terrains, they enjoy a much better PU because the driving force only needs to overcome friction, not support their own weight. To combine the advantages of both types of mobile robots, researchers propose various terrestrial-aerial vehicles [20, 3, 8, 17, 7, 23, 9, 10, 22, 2, 21, 11, 16, 14, 4, 6, 15, 18] (detailedly introduced in Sect.II-A). The locomotion modes of the vehicles can be dynamically altered depending on the need for better PU or higher mobility.
In this work, we customize a compact yet fully functional Two-wheeled terrestrIal-aErial quadrotor called TIE (as shown in Fig.1). Compared with the prototypes developed in previous works that share a similar structure [20, 9, 3, 8, 23, 17, 7], TIE possesses a smaller size, lighter weight, and longer hover duration. We also deploy adequate sensing and computing resources on TIE. Furthermore, careful engineering considerations are incorporated to balance its portability and robustness. The mobility of TIE is demonstrated in the video, in which it performs challenging tasks such as seamless locomotion transition, slope climbing, wall climbing, and rolling on rough ground.
The other focus of our work is autonomous navigation for terrestrial-aerial vehicles. At present, only Fan et al. present a complete terrestrial-aerial navigation framework. However, it is deficient in motion planning and motion control. The planning method proposed by Fan et al. cannot generate satisfactory trajectories due to the coarse path searching and the lack of trajectory optimization. Besides, it does not consider the nonholonomic constraint for terrestrial locomotion. As for motion control, the method of Fan’s fixes the total thrust in terrestrial locomotion, limiting PU improvement. We effectively solve these problems in the proposed navigation framework. Firstly, we develop a motion planner that generates safe, smooth, and dynamically feasible terrestrial-aerial trajectories. The nonholonomic constraint is also properly handled. Then, we design a unified terrestrial-aerial controller which includes an adaptive thrust adjustment method to improve PU in terrestrial locomotion. The results show up to 7 times less energy consumption compared with aerial locomotion. Furthermore, we integrate self-localization and local map fusion modules in the proposed navigation framework for real-world applications.
We perform sufficient experiments in challenging real-world environments to show the performance and robustness of the proposed robot system. During the tests, TIE plans safe and low-power trajectories in unstructured dense environments and accurately tracks these trajectories even when there are sharp turnings. We also compare our work with cutting-edge works. The results show that the proposed methods are superior in planning performance, controlling accuracy, and PU. Contributions of this letter are:
A lightweight and maneuverable quadrotor prototype that can seamlessly switch between terrestrial and aerial locomotion, and can carry sufficient onboard sensors and computing power.
An adaptive navigation framework that enables the proposed quadrotor prototype to traverse unknown cluttered environments with terrestrial-aerial integrated locomotion autonomously.
Presenting adequate real-world experiments and benchmark comparisons to validate the proposed robot system, as well as releasing source code and hardware configuration for the reference of the community.
Ii Related Work
Ii-a Terrestrial-Aerial Vehicle Design
Previous terrestrial-aerial vehicle designs can be divided into two categories according to whether the terrestrial locomotion is active or passive. Typically, the active ones use motor-driven wheels [22, 2, 11, 21] for ground locomotion. Because the dynamics and control of differentially driven wheeled mobile robots have been thoroughly studied, such design ensures the stability of terrestrial locomotion. Some other works include some novel deformable mechanisms [16, 14, 4] for locomotion mode switching. However, the above terrestrial-aerial vehicle designs need additional actuation and control systems for terrestrial locomotion or mode switching, which increase the vehicles’ payload and make the mechanical structure more complicated.
In contrast, the passive ones use wheels [20, 8, 3, 7, 17, 12, 1], cylindrical cages [9, 23] or spherical shells [6, 15, 18] that are driven by the multirotor thrust for terrestrial locomotion without introducing extra actuators or mechanisms. Among them, the cylindrical cage brings slip friction much bigger than the rolling one when turning, making it difficult to control the turning motion. The spherical shell makes the multirotor capable of omnidirectional rolling. However, this mechanism increases the instability of terrestrial locomotion, resulting in difficulties in the design of the control system. Also, the spokes of cages and shells block the view of the onboard cameras, which is hazardous for autonomous navigation. In brief, we can safely claim that the passive-wheeled quadrotor design is best suited for developing autonomous navigation systems.
Ii-B Autonomous Navigation for Terrestrial-Aerial Vehicles
As stated before, only a few researchers have worked on autonomous navigation for terrestrial-aerial vehicles. To the best of our knowledge, only Fan et al.  involve terrestrial-aerial motion planning. Firstly, it uses A* to search for a geometric path. Then, by adding a cost to nodes in the air, this method tends to search for a terrestrial path. Finally, a waypoint is selected along this path as the goal for a primitive-based local planner. It generates a set of minimum-snap trajectories, and scores each with a predefined cost to choose the best one. However, the path searching method is too coarse due to the lack of kinodynamic information. Also, since no post-refinement is applied to the trajectory in the local planner, its smoothness and dynamic feasibility cannot be guaranteed. Moreover, it does not consider the nonholonomic constraint in terrestrial locomotion.
|Reference||Size (mm)||Weight (g)||Hover Duration (min)||Onboard Sensing||Onboard Computing|
|Takahashi et al. |
|Kalantari et al. ||7|
|Colmenares et al. ||350|
|Nakao et al.||6400|
|Yamada et al. ||1360||8|
|Hada et al. ||1200||7|
|Fan et al. |
On the other hand, several works [7, 3, 20, 1] present control systems for passive-wheeled terrestrial-aerial quadrotor designs. Fan et al. and Colmenares et al. propose cascaded control schemes similar to a general quadrotor controller. They both simply set the thrust as a constant value lower than the vehicle weight, so the PU cannot be improved dynamically. In fact, the total thrust can be flexibly adjusted because the ground support force partially shares the vehicle’s weight. Takahashi et al.
propose a controller based on Linear Quadratic Regulator (LQR) with online parameter estimation. Nevertheless, no real-world trajectory tracking experiments are presented to validate the method’s efficacy. Atay et al. extend the works of [9, 20] by elaborating the specific dynamic model and developing a model-based control system. In addition, a thrust-optimization method is proposed. However, this work takes the pitch angle as one of the flat outputs of the controller, but fails to present the mapping from a given trajectory to the flat outputs, making this control system inapplicable to trajectory tracking.
Iii System Overview
Iii-a Robot Design
The main body of TIE is designed as a micro quadrotor with a diagonal wheelbase of 200mm. For terrestrial locomotion, we connect each passive wheel to a shaft fixed on the quadrotor, so that each wheel can rotate freely relative to the quadrotor. For strength and weight considerations, we use carbon fiber as the main structure of TIE, including the quadrotor frame, shafts, and wheels. Moreover, each passive wheel is made of two thin carbon plates fixed by nylon columns rather than a whole carbon plate, which further reduces weight while maintaining strength. The wheels, bearings, and shafts weigh 140g. The overall weight of the robot is 847.7g, including a 2300 mAh - 14.8 V battery that weighs 235g. It can hover up to 9 minutes in aerial locomotion with racing-drone brushless motors (T-Motor F2203.5 2850KV) and Gemfan 4023 propellers.
For autonomous navigation, we equip TIE with the following equipments onboard:
RealSense D430 depth camera 111https://store.intelrealsense.com/buy-intel-realsense-depth-module-d430-and-d4-board-bundle.html: This camera provides the depth images for local map fusion.
RealSense T261 tracking camera222https://store.intelrealsense.com/buy-intel-realsense-tracking-module-t261.html: This camera provides robust Visual Inertial Odometry (VIO) for UAV state estimation.
CUAV V5+ autopilot:333http://doc.cuav.net/flight-controller/v5-autopilot/en/v5+.html: It provides onboard IMU measurements and serves as the inner-loop controller.
Jetson Xavier NX444https://developer.nvidia.com/embedded/jetson-xavier-nx: It is an onboard computer with 6-core NVIDIA Carmel CPU and 8GB RAM. The entire pipeline, including map fusion, state estimation, motion planning and control modules, runs on it.
To make the electrical connections more integrated and stable, we also customize two printed circuit boards (PCBs) to connect the motors, electronic speed controllers (ESCs) and the above equipments. Detailed composition of TIE is shown in the exploded diagram Fig.3.
As mentioned above, several previous works develop terrestrial-aerial vehicles with a similar structure to the proposed one [20, 9, 3, 8, 23, 17, 7]. We summarize and compare hardware specifications from these works in Tab.I. It is worth noting that the TIE has the smallest size, the longest endurance, and a relatively light weight. On the other hand, only TIE and the one developed by Fan et al.  carry both onboard sensing and computing resources, while TIE weighs only as much as the other.
Iii-B Software Architecture
The architecture of the proposed navigation framework is illustrated in Fig.2
. Firstly, Visual Inertial Odometry (VIO) is obtained from the tracking camera. We fuse it with the IMU onboard the autopilot by Extended Kalman Filter (EKF) to generate smoother UAV state estimation. On the other hand, the depth images from the depth camera are projected to the world frame as a point cloud. We then adopt a column-wise evaluation to extract ground points. When maintaining an occupancy grid map, these points are not used, in order to avoid situations that the flat ground is set to be occupied. We also compute and update a Euclidean Signed Distance Field (ESDF) by an efficient algorithm developed by Zhou et al. . Afterward, the local planner searches for a kinodynamic path using the Fused VIO and the occupancy map. The resulted path is then optimized utilizing the gradient information obtained from the ESDF. The controller finally tracks the desired trajectory with both terrestrial and aerial locomotion. The experiments shown in Sect.VI-B validate the real-time performance of the proposed navigation framework.
Iv Safe Terrestrial-Aerial Motion Planning
The proposed terrestrial-aerial motion planner is built on Fast-Planner, which consists of a kinodynamic path searching method and a gradient-based spline optimizer. The path searching method is based on hybrid-state A* algorithm, which uses motion primitives instead of straight lines as graph edge in the searching loop. This work adds an extra energy consumption cost to the motion primitives whose destination is above the ground. Consequently, the path searching tends to plan terrestrial trajectories unless TIE encounters enormous obstacles and needs to fly over them, as shown in Fig.4.
In trajectory optimization, we reparameterize the generated trajectory as a degree uniform B-spine with control points
. Note that in terrestrial locomotion, we assume that TIE moves over flat ground, so that the vertical motion can be omitted. We then classify the control points above the ground as, and the rest as . Each terrestrial control point is two-dimensional, i.e., . To refine the trajectory, we firstly adopt the following cost terms from :
where is the smoothness cost designed as an elastic band cost function. is the collision cost based on the ESDF gradient information. and are dynamical feasibility costs that limit velocity and acceleration. are weights for each cost terms. Due to the convex hull property of the B-spline, the above cost terms only constrain the control points for safety and dynamical feasibility. We refer the readers to  for detailed formulations.
In terrestrial locomotion, the velocity of TIE is limited to be parallel with the yaw angle due to the nonholonomic constraint. Therefore, if the trajectory is too curved, huge tracking errors will occur during turning. To resolve this, we enforce a cost on to limit the curvature of the terrestrial trajectory. The curvature at is defined as , where , and . Therefore, this cost can be formulated as
where is a differentiable cost function with specifying the curvature threshold:
The derivation of the gradient can be found in . Note that may be segmented into several subsets by intermediate aerial control points, the curvature of the endpoints are not taken into consideration. In general, the overall objective function is formulated as follows:
The optimization problem is solved by a non-linear optimization solver NLopt555https://nlopt.readthedocs.io/en/latest/.
After motion planning is done, a setpoint on the generated trajectory is selected according to the current timestamp, and then sent to the controller as a reference state in the inertial frame (defined in Fig.5). An aerial setpoint includes the yaw angle and 3D position, velocity, and acceleration (). A terrestrial one includes the yaw angle and 2D position and velocity (). For consistency, and are both set to be parallel with the velocity. If the current setpoint is in a different locomotion mode than the previous one, an extra trigger will be sent to the controller for the locomotion switch.
V Unified Terrestrial-Aerial Motion Control
This section manifests the proposed controller, which adopts a cascaded architecture for both terrestrial and aerial locomotion. The reference frames are defined in Fig.5. The estimated state obtained from onboard VIO is denoted as , including the position, velocity, orientation (parameterized by Euler angles ) and its derivation (). As for the locomotion switch, When take-off is desired, the controller immediately switches to aerial mode without a slow transition process. During landing, the controller commands TIE to land smoothly with a constant speed, avoiding a sudden impact that may cause VIO divergence.
V-a Aerial Controller
The aerial controller is shown in Fig.6. It takes the reference state from motion planning as the input. The position control module computes the position and velocity error using a proportional controller, and combines them with the reference term to generate a desired acceleration . is firstly used to generate the desired thrust , and then used together with to calculate the desired attitude leveraging the differential-flatness of quadrotors. The detailed equations of attitude calculation can be found in . The inner attitude control and body-rate control generate the attitude derivations
and the desired moment, respectively. These two modules are performed on the onboard pilot using PX4 open-source firmware666https://github.com/PX4/PX4-Autopilot.
V-B Terrestrial Controller
Fig.7 illustrates the terrestrial controller, which owns a similar architecture to the aerial one. The attitude controller is executed by the onboard autopilot as well. The terrestrial controller’s tracking performance is shown in Fig.8.
1) Yaw Control: The desired yaw is calculated according to the current position error between and , which is defined as . When the norm of is relatively small, is taken as the reference term which points along the trajectory’s tangent direction. However, if the norm is larger than a threshold, is calculated to be parallel with for error correction. The corresponding equations are shown as follows:
where is the position error threshold. and are the x-axis and y-axis value of , respectively.
2) Adaptive Thrust Control: The desired total thrust is adaptively controlled according to the magnitude of current desired turning angle, defined as . As mentioned before, position tracking error accumulates when TIE is turning due to the nonholonomic constraint. To reduce the tracking error, we dynamically adjust the thrust so that it produces a maximal yaw acceleration large enough to make TIE finish the turning in a short period , which is set as in experiments. Since is small, almost remains constant, and the yaw kinematics can be derived:
Due to the nonlinearity of the inner attitude controller, we obtain the relationship between and by experimental fitting, given as
3) Attitude Calculation:
It generates the desired attitude . We firstly calculate the desired attitude in inertial frame and obtain by coordinate transformation. Among them, has been computed in the yaw controller, and remains zero due to the assumption that TIE moves on flat ground. The following equations give the derivation of . Firstly, we compute the x-axis value of the desired acceleration in terrestrial frame (denoted as ) based on and the velocity error with a feedback control law:
where is the integral velocity tracking error. , and are constant gains. Then, can be calculated with the following dynamics equation. In this work, we do not take into account external forces such as the rolling friction.
where is the total mass of TIE, and is the scale parameter.
|Inte. of Acc.||
Vi-a Benchmark Comparisons
To demonstrate the superiority of the proposed navigation framework, we conduct benchmark comparisons against the previous works on terrestrial-aerial navigation in two-folds: the terrestrial-aerial planning and the terrestrial controller.
1) Comparison of Terrestrial-Aerial Planning: We conduct comparisons between the proposed planning method and Fan’s . Specifically, each algorithm runs for 50 times independently in a simulation environment with 80 randomly deployed obstacles. We only compare planning methods and do not include terrestrial-aerial motion controllers in the simulation tests. The distance between the starting and goal positions is . We also set up a huge barricade between the starting and goal positions, requiring the robot to fly over it. All the computations are done on a 2.9 GHz processor with 16 GB RAM. The velocity and acceleration limits are set as and . As shown in Tab.II, the proposed planner finds trajectories with less computing time, better smoothness (integral of the squared acceleration), and higher success rate. Firstly, our planner both refines the trajectories’ smoothness and dynamical feasibility, which are not considered in Fan’s . In addition, the motion-primitive based method in Fan’s  is time-consuming and incomplete, which may fail to generate feasible trajectories when facing complex environments, resulting in a low success rate even with a higher computing time. Therefore, our method outperforms Fan’s  in both time efficiency and planning performance.
2) Comparison of Terrestrial Controller: We compare the proposed terrestrial controller with method[3, 7] in real-world environments. Only the outer translation control is compared because it is our focus, and the inner attitude control of all methods is executed by the autopilot. During the comparison, TIE uses each controller to track a lemniscate trajectory with different velocities. Since method[3, 7] set the desired thrust to be constant, we first test our method, then calculate the desired average normalized thrust (denoted as ) and assign it to method[3, 7]. The average and maximal trajectory tracking error and are compared. As shown in Tab.III and Fig.9, the proposed method achieves lower and with the same in every case. With the adaptive thrust control, the proposed controller generates a larger thrust to make TIE pass through sharp turnings faster, thereby reducing the tracking error. In contrast, method[3, 7] requires a large thrust at all times to achieve the same tracking performance, which greatly increases energy consumption.
To demonstrate the autonomy and performance of the entire robot system, we perform extensive autonomous tests in various complex environments (as shown in Fig.11). Except for several waypoints, no prior information of the environments is given. The unknown dense environments and limited onboard vision make the experiments challenging.
1) Walking out of a Complex Maze:
In this experiment, TIE has to navigate a complex maze with tremendous obstacles, sharp turnings, and an unavoidable barricade. The velocity limit is set to be . It turns out that TIE manages to walk out of this maze, and it remains in terrestrial locomotion except when it flies over the unavoidable barricade. This result is as expected because terrestrial locomotion is preferable due to better PU.
2) Flying vs Rolling: This experiment presents quantitative PU comparisons between terrestrial and aerial locomotion. The experimental scene is complicated as well because of tortuous paths and a great many obstacles. However, it is set to be passable for both terrestrial and aerial locomotion. In the experiment, TIE passes through the environment in terrestrial and aerial locomotion with the same velocity limit , respectively. The normalized thrust curve is depicted in Fig.12. The average thrust is in terrestrial locomotion and in aerial locomotion. That is, TIE passes this challenging test in both locomotion modes, but requires about only a quarter as much thrust in terrestrial locomotion as in the aerial one. We also experimentally measure that the corresponding energy consumption ratio is approximately . This result highlights the great advantage of terrestrial locomotion in PU.
3) Moving through Winding Tunnels: We perform aggressive flight and rolling tests in this experiment to present the proposed system’s high mobility even in autonomous navigation. Two winding tunnels are set for the flight and rolling test, respectively. The end of each tunnel is outside TIE’s sensing range, so TIE needs to replan in time and turn quickly to pass the test. As a result, TIE can travel back and forth through the tunnels with a velocity up to in aerial locomotion and in terrestrial locomotion. The results are comparable with state-of-the-art autonomous quadrotor systems in [25, 26, 24].
4) Traversing a large Office: The last experiment is conducted in an unknown office with a size over . It is full of cluttered objects, leaving only narrow passages, which brings difficulties to motion planning. What is more, the lighting and terrain condition around the office does not remain the same, posing a huge challenge to the perception and the terrestrial controller of the system. In order to test both terrestrial and aerial navigation, We set up a take-off waypoint halfway and keeps TIE flying after it passes this waypoint. It turns out that TIE safely traverses the office in terrestrial-aerial integrated locomotion. The executed trajectories are shown in Fig.10. This experiment strongly demonstrates the robustness of the proposed system.
In conclusion, the above experiments verify the excellent performance of the proposed robot system from all aspects. It maintains the high maneuverability of a quadrotor while incorporating low-power terrestrial locomotion and can flexibly switch locomotion modes according to environmental changes. Moreover, the system’s robustness makes it capable of both aggressive locomotion and long-time navigation even in unknown dense environments. More details are included in the video.
Terrestrial-aerial vehicles possesses distinct advantages because they combine both the mobility of UAVs and the long endurance of UGVs. However, there are no representative works on the autonomous navigation of these vehicles. To fill this gap, we present an autonomous quadrotor system that can safely navigate in unknown cluttered environments with terrestrial-aerial hybrid locomotion. Based on the design of a lightweight terrestrial-aerial quadrotor, we propose an adaptive navigation framework which mainly consists of a safe motion planner and a unified motion controller. We carry out challenging tests to show the proposed system’s robustness and superiority.
For future work, we will pay attention to motion planning on uneven terrain. On the other hand, we will consider the exogenous forces in the controller, so as to further improve the control accuracy and energy savings.
-  (2021) Control and control allocation for bimodal, rotary wing, rolling–flying vehicles. Journal of Mechanisms and Robotics 13. Cited by: §II-A, §II-B.
-  (2021) BAXTER: bi-modal aerial-terrestrial hybrid vehicle for long-endurance versatile mobility: preprint version. arXiv preprint arXiv:2102.02942. Cited by: §I, §II-A.
-  (2019) Nonlinear control for ground-air trajectory tracking by a hybrid vehicle: theory and experiments. IFAC-PapersOnLine 52 (8), pp. 19–24. Cited by: §I, §I, §II-A, §II-B, TABLE I, §III-A, §VI-A, TABLE III.
-  (2021) Design and analysis of fcstar, a hybrid flying and climbing sprawl tuned robot. IEEE Robotics and Automation Letters. Cited by: §I, §II-A.
-  (2008) Practical search techniques in path planning for autonomous driving. Ann Arbor 1001 (48105), pp. 18–80. Cited by: §IV, §IV.
-  (2015) A micro spherical rolling and flying robot. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5863–5869. Cited by: §I, §II-A.
-  (2019) Autonomous hybrid ground/aerial mobility in unknown environments. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3070–3077. Cited by: §I, §I, §I, §II-A, §II-B, §II-B, TABLE I, §III-A, §VI-A, §VI-A, TABLE II, TABLE III.
-  (2017) Development of a bridge inspection support system using two-wheeled multicopter and 3d modeling technology. Journal of Disaster Research 12 (3), pp. 593–606. Cited by: §I, §I, §II-A, TABLE I, §III-A.
-  (2013) Design and experimental validation of hytaq, a hybrid terrestrial and aerial quadrotor. In 2013 IEEE International Conference on Robotics and Automation, pp. 4445–4450. Cited by: §I, §I, §II-A, §II-B, TABLE I, §III-A.
-  (2014) Modeling and performance assessment of the hytaq, a hybrid terrestrial/aerial quadrotor. IEEE Transactions on Robotics 30 (5), pp. 1278–1285. Cited by: §I.
-  (2020) Drivocopter: a concept hybrid aerial/ground vehicle for long-endurance mobility. In 2020 IEEE Aerospace Conference, pp. 1–10. Cited by: §I, §II-A.
-  (2021) Driving and tilt-hovering–an agile and manoeuvrable aerial vehicle with tiltable rotors. IET Cyber-Systems and Robotics. Cited by: §II-A.
-  (2011) Minimum snap trajectory generation and control for quadrotors. In 2011 IEEE international conference on robotics and automation, pp. 2520–2525. Cited by: §V-A.
-  (2018) A multi-modal hovering and terrestrial robot with adaptive morphology. In Proceedings of the 2nd International Symposium on Aerial Robotics, Cited by: §I, §II-A.
-  (2015) Proposal and experimental validation of a design strategy for a uav with a passive rotating spherical shell. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1271–1278. Cited by: §I, §II-A.
-  (2017) A small hybrid ground-air vehicle concept. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5149–5154. Cited by: §I, §II-A.
-  (2019) Development of a bridge inspection support robot system using two-wheeled multicopters. Journal of Robotics and Mechatronics 31 (6), pp. 837–844. Cited by: §I, §I, §II-A, TABLE I, §III-A.
-  (2019) Rollocopter: an energy-aware hybrid aerial-ground mobility for extreme terrains. In 2019 IEEE Aerospace Conference, pp. 1–8. Cited by: §I, §II-A.
-  (2018) Lego-loam: lightweight and ground-optimized lidar odometry and mapping on variable terrain. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4758–4765. Cited by: §III-B.
-  (2015) All-round two-wheeled quadrotor helicopters with protect-frames for air–land–sea vehicle (controller design and automatic charging equipment). Advanced Robotics 29 (1), pp. 69–87. Cited by: §I, §I, §II-A, §II-B, TABLE I, §III-A.
-  (2021) Multimodal dynamics analysis and control for amphibious fly-drive vehicle. IEEE/ASME Transactions on Mechatronics 26 (2), pp. 621–632. Cited by: §I, §II-A.
-  (2017) A design of a small mobile robot with a hybrid locomotion mechanism of wheels and multi-rotors. In 2017 IEEE International Conference on Mechatronics and Automation (ICMA), pp. 1503–1508. Cited by: §I, §II-A.
-  (2017) Development and field test of novel two-wheeled uav for bridge inspections. In 2017 International Conference on Unmanned Aircraft Systems (ICUAS), pp. 1014–1021. Cited by: §I, §I, §II-A, TABLE I, §III-A.
-  (2020) Tgk-planner: an efficient topology guided kinodynamic planner for autonomous quadrotors. IEEE Robotics and Automation Letters 6 (2), pp. 494–501. Cited by: §I, §VI-B.
-  (2019) Robust and efficient quadrotor trajectory generation for fast autonomous flight. IEEE Robotics and Automation Letters 4 (4), pp. 3529–3536. Cited by: §I, §III-B, §IV, §IV, §VI-B.
-  (2021) EGO-planner: an esdf-free gradient-based local planner for quadrotors. IEEE Robotics and Automation Letters 6 (2), pp. 478–485. External Links: Cited by: §I, §VI-B.