Modular and Scalable Fabrication of Insect‐Scale Aerial Robots toward Demonstrating Swarm Flights

Insects can navigate in cluttered spaces and perform challenging functions such as pollination and collective object transport. By exploiting scaling laws and bioinspired designs, insect‐scale micro‐aerial‐vehicles (MAVs) have demonstrated impressive flight capabilities such as in‐flight collision resilience and acrobatic maneuvers. However, existing subgram MAVs are difficult to design, construct, and repair. Coupled with challenges in robot sensing and control, existing subgram MAVs have not achieved insect‐like swarm flight, which limits potential studies of swarm behaviors and future applications such as collective sensing. Herein, a new design and fabrication method is developed to substantially improve the fabrication scalability of subgram MAVs. Based on a small set of design parameters, an automated algorithm generates the laser cut files of microrobotic components. To reduce fabrication and assembly time, stereolithographic 3D printing is used for making static components such as the airframe and connectors. The modular design enables straightforward assembly and repair, which reduces the overall fabrication time by over 2 times. Owing to the ease of fabrication and good reliability, two subgram MAVs demonstrate controlled hovering flight and coordinated lifting of an object. This result lays the foundation for future robotic studies of collective insect flight.


Introduction
Honeybees are agile pollinators that navigate through cluttered environments, land on delicate flowers, and construct complex structures such as beehives. [1,2]Owing to their small scale (<1 g, <5 cm), aerial insects develop several unique advantages compared to mesoscale (>5 g, >10 cm) micro-aerial-vehicles (MAVs).For instance, due to the diminishing inertial effects, aerial insects are resilient against in-flight collisions [3] and can perform acrobatic body flips within 40 ms. [4]They can also leverage the dominating surface effects to perch and land on inclined and compliant surfaces. [4]Developing insect-scale aerial robots not only offers a platform to investigate unique biomechanics and aerodynamics, but also inspires future applications in environmental monitoring and assisted agriculture.
[8] These state-of-the-art MAVs have demonstrated impressive flight capabilities such as hovering, [6] trajectory following, [9,10] and perching. [11][14][15] However, due to limitations in robustness and fabrication scalability, these subgram MAVs have not yet demonstrated simultaneous flights.Unlike aerial insects, most subgram MAVs are fragile against in-flight collisions, which prohibit them from flying near obstacles or each other.Furthermore, the existing design and fabrication methods have not been optimized for scalable production and repair of subgram MAVs.
To improve the robustness of subgram MAVs, our team developed a new class of soft-actuated aerial robots. [16]Known as SoftFly, the robot is driven by a dielectric elastomer actuator (DEA), which demonstrates muscle-like power density and resilience.In addition to achieving controlled flights, [16,17] our prior works have shown insect-like flight resilience, which paves the way for enabling swarm flight.However, the robot design and fabrication process remained tedious and error prone, where a researcher needed to manually design all laser cut files and later assemble the robot from 2D laminate structures.In the design process, it remained difficult to vary robot kinematic parameters such as the transmission ratio and airframe size.The Smart Composite Manufacturing [18,19] (SCM) process constructs 3D components out of many layers of 2D laminate materials.A simple change of one design parameter requires modifications of all corresponding laminate layers, which is a major challenge for fast design iterations.In the SCM fabrication process, the assembly of precise robotic components requires delicate manual operation under a confocal microscope.Permanent epoxies are applied in the assembly process, and they prevent replacement and repair of key robot components such as transmissions and actuators. [20]These limitations of robot design and fabrication further restrict researchers from achieving insect-like simultaneous flights.
Several new design methods were developed to address the abovementioned challenges.A software called popupCAD simplifies the design process by automatically updating the dependency relationships between different material layers. [21,22]To simplify manual assembly, a monolithic design approach was proposed in which most robot components are made from a single laminate. [7]The robot is folded into a 3D structure, and this process reduces assembly time while improving the overall precision.However, these methods still require substantial expertise and overlook key issues including fabrication scalability and repairability.A popupCAD user needs to define dependency relationships for each robot component, and manually organize component placements in preparation for laser micromachining.To design a monolithic robot in a single laminate, a researcher needs to carefully consider spatial interference of different material layers and robotic components.This fabrication process mostly focuses on producing a single robot, and it does not consider repairability when a flexure or structural component is damaged.Consequently, despite having these novel design approaches, there still lacks scalable and reliable subgram MAVs that can fly together in a swarm.
Herein, we develop novel design and fabrication methods for scalable production of subgram MAVs.We simplify the robot design process by identifying key robot parameters such as the transmission ratio, wing hinge stiffness, and wing size.Based on user-defined parameters, our design algorithm automatically generates cut files for laser micromachining.Compared to prior methods, [23] our approach reduces the design time from hours to within seconds.To accelerate the fabrication process, we classify microrobotic components into two groups: static structural support parts and high precision (<10 μm) kinematic parts.For structural support parts that do not require high precision, we used stereolithographic 3D printing to automate the fabrication process.For kinematic parts that have precise transmission ratio and stiffness, we fabricate them in a large number (>40) through the SCM method.Our design and fabrication approach not only reduces the fabrication and assembly time by over an order of magnitude, but also substantially improves robot consistency and repairability.Based on this new approach, we further upgrade our flight controller and motion capture system to enable simultaneous flight.We demonstrate a 10 s hovering flight performed by 2 subgram MAVs (Figure 1A).In comparison, the new robot only requires 50% time to fabricate while maintaining similar lift forces and endurance.Furthermore, these two MAVs can collaboratively carry an irregularly shaped payload (Figure 1B).These flight results represent important progress for subgram MAVs, and they lay the foundation for future studies of insect-scale collective flight behaviors.

Robot Design and Fabrication
To achieve scalable microrobot fabrication, we adopt a modular design where each robot consists of four identical units (Figure 2A).Each module is independently controlled during flight, and it can be easily replaced without influencing the functions of neighboring ones.In contrast to prior works [7] that aimed to reduce the number of discrete robot components, we focus on improving robot repairability in the design process.illustrates a robot module that consists of ten components: an airframe, a soft actuator, two connectors, two linear four-bar transmissions, two wing hinges, and two wings.To improve repairability, we only apply UV-cured epoxy (UV Liquid, Bondic) in the assembly process, where each component can be easily removed and replaced.An advantage of this design is that each discrete robot component can be produced in a large quantity, and due to the ease of assembly, many robot modules can be assembled and repaired efficiently.
To further automate the design and fabrication process, we classify robot components into two classes: static structural parts and kinematic parts that are crucial for robot flight.The static structural parts consist of the robot airframe and connectors (Figure 2C-F), which do not require high precision.We use stereolithographic 3D printing to fabricate these robot parts that have a feature size of approximately 100 μm.In a single-batch fabrication, our 3D printer (Form 3, Formlabs) can make a sufficient number of components for assembling eight robots (Figure 2C).In contrast, the robot actuator, kinematic components, and wings require a finer feature size (%10 μm).The dielectric elastomer actuator is made using an existing method, [23] and the robot transmissions, wing hinges, and wings are made in large quantities using the SCM process.To improve the design efficiency, we develop a parametrized algorithm to automatically generate the laser cut files as required by the SCM process.In the next two sections, we describe the 3D printing and the SCM fabrication processes in detail.Table 1 summarizes the robot design and fabrication time.A detailed comparison of prior and proposed fabrication methods is described in Notes S1, Supporting Information.

Stereolithographic 3D Printing of Static Components
We choose commercially available stereolithographic 3D printing methods for fabricating static robot components including the robot airframe and connectors.In prior works, [16,23] the robot airframe was made of eight pieces of planar carbon fiber laminates that were manually assembled into a 3D structure.The airframe assembly time is approximately an hour, and then a researcher manually applies permanent glue (495, Loctite) along all joints (see Notes S1 and Figure S1, Supporting Information, for details).The carbon fiber airframe weighs 30-32 mg.To minimize airframe weight while maintaining structural stiffness, we choose a high-modulus material (Rigid 10 K Resin, Formlabs).The 3D-printed airframe (Figure 2E) weighs 33-34 mg and requires roughly an hour of printing time.Similarly, the new connectors between the actuator and the transmission are also 3D printed (Figure 2D).Fabrication automation of these components substantially reduces the required manual assembly time (Table 1) without sacrificing the robot flight performance.
In this work, we also 3D-print the structure that connects all four robot modules (Figure 2F).Over a longer distance (3-5 cm), the 3D-printed support beams require 250% larger crosssectional area to achieve a similar structural stiffness.Compared to the 56 mg carbon fiber connectors in the prior work, the 108 mg 3D printed connector is 92% heavier, which reduces the net robot payload.Despite this payload reduction, the 3Dprinted connector substantially reduces the robot assembly time (Table 1).To connect four independent robot modules (Figure 1) with the 3D-printed connector, it only requires approximately 10 min (Video S1, Supporting Information), whereas the carbon  fiber connectors require an hour (Figure S3, Supporting Information, in ref. [23]).

Automated Generation of Laser Cut Files for the SCM Process
The robot kinematic components refer to the parts that either deform or move relative to the robot center of mass.For a robot to achieve precise and well-controlled flapping kinematics, these parts must have a high precision of approximately 10 μm.Prior works [18,19] developed the SCM fabrication method where 2D planar materials were laser micromachined and then laminated into a 3D structure.We adopt the SCM method to fabricate the wing hinges, transmissions, and wings that move at 400 Hz.One challenge of the SCM fabrication approach is that a robot component consists of multiple (3-20) material layers, where the laser cut file of each layer must be individually generated.To simplify this design process, we implement an algorithm that automates laser cut file generation based on key kinematic parameters.The robot components are produced in a large number to improve fabrication scalability and consistency.In the next sections, we describe the design and fabrication of the wing hinges, transmissions, and wings.

Wing Hinge
The robot wing hinge is a kinematic component that connects the transmission and the wing (Figure 2B).During flight, the wing hinge acts as a torsional spring that passively mediates the wing pitch rotation.We adopt the SCM method for batch design and fabrication of the wing hinge.Figure 3A illustrates a design cut file where 60 wing hinges are arranged in a 29 mm Â 29 mm template.Figure 3B shows the corresponding laminate material after all wing hinges are extracted.
Figure 3C shows the top and side view images of a wing hinge.In this work, the wing hinge consists of nine material layers: three polyimide flexure layers, two carbon fiber structural layers, and four adhesive layers.The central polyimide layer acts as a rotational joint and the outer layers serve as linear torsional springs.The key design parameters h w and h l are the hinge width and length, respectively (Figure 3C).These design parameters determine the wing hinge stiffness (K ) through Equation ( 1) where E is the Young's modulus of the compliant polyimide material and h t is the polyimide material thickness.In this work, the central and outer polyimide (Kapton) material thickness are chosen as 7.5 and 12.7 μm, respectively.An important fabrication challenge arises from the 5% material thickness variation in commercially available polyimide materials.To precisely control the wing hinge stiffness, we modify the values of h w and h l based on the measured material thickness.This condition requires us to update the laser cut files each time a new batch of wing hinges is made.
Figure 3D shows the laser cut files of the nine material layers and the final laser release cut.Traditionally, modifying h w and h l is a tedious design process where a researcher needs to manually update all the design files.In this work, we develop a parametrized design code that automatically generates all the design files based on h w and h l .The code calculates the geometric dependence between each material layer, creates the 10 design files, and then repeats this process for all 60 wing hinges.The code also tiles the wing hinges within the 29 mm Â 29 mm template.The parameters h w and h l of each wing hinge can be individually updated.With this code, the laser cut file generation time reduces from hours to within seconds (Table 1).A researcher only needs to prescribe key kinematic parameters while the tedious work of updating cut files is automatically done by the code.The fabrication process still requires manual preparation of different materials and supervision of laser micromachining.On average, it requires approximately 2 h for making one batch of 60 wing hinges, which is sufficient for assembling 30 robot modules (see Notes S2 and Figure S2, Supporting Information, for details).

Four-Bar transmission
The robot transmission connects the actuator, the airframe, and the wing hinge.Based on a linear four-bar mechanism, the transmission converts the actuator's translational motion into a rotational wing stroke motion.We follow a prior work's design [24] where 48 transmissions are tiled within a 29 mm Â 29 mm template.In contrast to designing and tiling cut files manually, our code automatically calculates all laser cut files based on user-defined parameters.Figure 3E,F shows the tiled design file and the corresponding laminated structure after all transmissions are extracted.Robot transmission has two key design parameters: transmission stiffness k t and transmission ratio r t .The transmission stiffness k t determines the robot resonant flapping frequency, [16] and it can be adjusted through the transmission height t h (Figure 3G).The transmission consists of 13 layers: 5 carbon fiber structural layers, 2 polyimide flexural layers, and 6 adhesive layers.As illustrated in Figure 3H, all cut files need to be updated accordingly when the parameter t h is modified.Our code automatically generates the cut files related to the 13 material layers and the final release cut.Similar to that of robot wing hinges, the overall design time is reduced to seconds (Table 1).This implementation substantially simplifies and accelerates the design process.
The second design parameter is the transmission ratio r t .As shown in a prior work, [25] r t can be approximated as 1=t l , where t l is the separation distance between the two flexural layers.In this design, t l is the total thickness of layers 4-10 and it is chosen to be 380 μm.To achieve this thickness, the central carbon fiber layer (CF 7) is used as a spacer.It is made based on the thickness measurement of all other material layers.Compared to another approach [6,7] in which the transmission is manually folded from a five-layered laminate, this method improves fabrication scalability and consistency.It requires approximately 3 h to produce 48 robot transmissions, which can assemble 24 robot modules.

Wing
A pair of robot wings generates the lift force during flight.The wing is 1 cm long, weighs 0.5 mg, and flaps at the operating frequency of 400 Hz.Our prior works [26,27] showed the robot's  net lift force and aerodynamic efficiency critically depend on wing shape and inertia.Through experiments, we find that our 3D printing method cannot fabricate wings of sufficient structural stiffness while maintaining low inertia.Hence, we fabricate the wings through the SCM method.Based on a prior work, [27] we improve an existing wing design algorithm to enable scalable wing fabrication.Figure 4A,B shows a design file and the corresponding laminate materials where 24 wings are made in a 43 mm Â 43 mm template.
Figure 4C shows a wing that is made of three material layers: carbon fiber structural support, adhesive, and polyester membrane.The wing cut files are specified by several parameters such as the wingspan, its first and second area moment, and spar width.Given these parameter values, our code automatically generates the laser cut files (Figure 4D) and fits 24 wings within the template.
Compared to our prior work, [27] this new wing design implementation substantially improves fabrication consistency.A key challenge of wing fabrication relates to the requirement of high precision (<20 μm error).The wing spar width varies between 80 and 110 μm, and a small variation (10%) can cause a substantial change of the system resonance frequency.While the laser micromachining process can ensure high consistency (<5 μm), material deformation and realignment during release cut can lead to large errors (%30 μm).Specifically, after the initial carbon fiber and adhesive layers are laser micromachined (Figure 4D), they are laminated with a polyester membrane under 200 °C for 2 h.In this process, the laminated materials slightly deform and expand by 10-20 μm.Next, the laminated material is placed under the laser, realigned, and then laser cut again to release the wings.The accumulated error in the lamination and realignment process can cause large inconsistencies of over 20 μm. Figure 4E shows two wings made with the prior method.The realignment errors cause different spar widths, which result in a 50 Hz change of system resonance frequency.In the past, every wing pair was manually selected and experimentally trimmed, and this was a time-consuming process.
In this work, we modify the carbon fiber cut file design for improving fabrication consistency.Figure 4F compares the prior and current designs where the carbon fiber and the release cut files are overlaid.In the prior design, a small realignment offset can change the leading-edge spar width.In the new design (Figure 4F), most of the release cut contour is also traced in the initial carbon fiber cut.The final release cut removes a few structural connections.The realignment error only causes small offsets in these connecting locations.With this new design, the manual wing pairing process is removed, and variation of flapping frequency is controlled within 10 Hz.All robot modules operate at 400 Hz without requiring wing pairing or trimming.

Robot Characterization and Flight Demonstrations
Based on the design and fabrication methods in Section 2, we construct a new robot and perform characterization and flight experiments.We further upgrade our flight arena to demonstrate simultaneous hovering flight of two robots.In the next sections, we describe the flight arena setup, robot characterization experiments, and flight demonstrations.

Experimental Setup
We expand an existing flight arena [23] to enable simultaneous flight of two robots.The flight arena (Figure 5A) consists of a motion capture system (Vantage V5, Vicon), a high-speed camera  [27] and new designs.The new cut file design substantially reduces realignment error and improves the consistency of wing inertia.
(Phantom VEO 710, Vision Research), eight high-voltage amplifiers (2220, Trek), safety switches, a host computer, and a realtime target computer (Baseline Target Machine, Speedgoat).In our prior work, [23] the tracking data was sent to the target machine through serial port communication.To enable multiobject tracking, we switch to Ethernet communication where the positions and attitudes of multiple objects can be tracked at 400 Hz.The new flight controller is written in MATLAB Simulink where it can control two robots at a feedback rate of 5 kHz.The six Vicon cameras provide a tracking volume of 1 m Â 0.8 m Â 0.6 m.

Robot Characterization
Prior to conducting controlled flight experiments, we characterize robot flapping-wing motion and lift force generation.
Figure 5B shows a composite image of the robot flapping-wing kinematics at 1600 V.The measured wing stroke and pitch kinematics (Figure 5C) have amplitudes of 37.5°and 63.7°, respectively.This wing kinematics is similar to that of the original robot (Figure 5C).Video S2, Supporting Information, also compares the two robots' wing motions.In addition, in preparation of flight experiments, we conduct liftoff tests to measure the robot net lift force.Figure 5D shows a composite image of robot liftoff where it is mounted on a liftoff stand.The robot carries a 100 mg payload and ascends upward.Using a prior tracking method, [23] we calculate the net lift force and repeat this experiment for different driving conditions.Figure 5E shows the measured lift force at four driving voltages, and we use these measurements to update the parameters in the flight controller.The liftoff experiment is also shown in Video S2, Supporting Information.Compared to the original robot module, [23] the new robot with a 3D-printed airframe and connectors exhibit similar performance.

A 5 s Feedback-Controlled Hovering Flight
Next, we evaluate the new robot's flying performance through a 5 s hovering flight.In this experiment, we command the robot to hover at 5 cm above ground for 5 s. Figure 6A shows a composite image of this flight, which consists of a 0.4 s takeoff, 4.2 s hovering, and 0.4 s landing.The measured lateral position, altitude, and attitude are shown in Figure 6B-D, respectively.During the hovering phase, the root mean square errors (RMSEs) of position, altitude, and attitude are 2.05, 0.30 cm, and 1.07°, respectively.In this flight (Figure 6A and Video S3, Supporting Information), we do not control the robot yaw motion.Compared to prior flights performed by the original robot, [23] this new robot with 3D-printed structural supports demonstrates similar flight accuracy.This result shows the new design and fabrication methods accelerate robot construction without compromising flight performance.

A 10 s Simultaneous Flight of Two Robots
Based on the 5 s hovering demonstration, we conduct a 10 s simultaneous flight of the two robots.Figure 7A shows a composite image of this flight, where the two robots are synchronously commanded to liftoff (0.6 s), hover (9 s), and land (0.4 s).The robot hovering setpoints are separated by 13 cm, making the tip-to-tip distance 6 cm-approximately 1 body length apart.Figure 7B-G shows the tracked positions, altitude, and attitude of each robot.For the original robot in its hovering phase, the RMSEs of position, altitude, and attitude are 1.18, 0.54 cm, and 0.87°, respectively.The corresponding errors of the new robot are 1.20, 0.15 cm, and 1.12°, respectively.As shown in Figure 7 and Video S4, Supporting Information, the original and the new robots show similar flight performance.The original robot has a 3.6 times larger altitude error (Figure 7C), which is likely caused by imperfect controller parameter tuning.This flight demonstration is a significant result for the field of insect-scale aerial robotics.It represents the first simultaneous flight performed by two subgram robots thanks to the scalable fabrication methods proposed in Section 2, and it highlights good robot consistency and controllability.

A 5 s Collective Payload Lifting Demonstration
We further perform a collective payload lifting demonstration where the two robots cooperatively carry a 60 mg and 13 cm long object during flight.Figure 8A shows a composite image of this flight that consists of 0.8 s ascent, 3.8 s hover, and 0.4 s descent.Although the payload only weighs 10% of the robot, it is difficult to be lifted by a single robot due to its geometry.To transport this object, we attach two strings to its two ends and then connect each string to a robot.The control parameters are tuned to accommodate the added weight of the payload, but the controller does not take into account the additional torques and lateral forces caused by the object.During this collective flight, the two robots maintain relatively accurate separations while carrying the object.Figure 8B-G shows the measured lateral positions, altitude, and attitude.For the original robot and the new robot, the RMSEs of lateral position are 1.27 and 1.60 cm, the altitude RMSEs are 0.17 and 0.23 cm, and the attitude RMSEs are 1.14°and 1.66°, respectively.Even though the lifted object create extra forces and torques to the robots, the RMSEs of lateral position and attitude in this flight only increase by 0.25 cm and 0.41°on average.Overall, Figure 8 and Video S5, Supporting Information, show a well-coordinated lifting demonstration-the very first coordinated aerial behavior at this scale.This unprecedented flight highlights good robot controllability, and it shows the promise of enabling collaborative swarm flight in future subgram aerial robots.

Conclusion
In this work, we develop parametrized design and efficient fabrication methods for constructing insect-scale aerial robots.In the design process, our code takes in user-defined parameters and automatically generates the laser cut files as required by the SCM fabrication method.This approach reduces the design time from hours to within a few seconds, which enables fast design iteration.In addition, we classify robot components into two classes: highly precise (<10 μm) kinematic components and structural support components.The kinematic components such as wings, wing hinges, and transmissions are made in a large number (24-60) through the SCM process.The structural components are fabricated through commercially available stereolithographic 3D printing.These novel approaches substantially improve fabrication scalability and consistency-reducing robot fabrication time by over 50% while maintaining similar performance.The modular design further improves robot repairability where any damaged component can be replaced easily.These design and fabrication improvements enable the first demonstration of simultaneous flight in subgram aerial robots.Two robots achieved a 10 s hovering flight under external motion tracking and feedback control.In addition, these two robots demonstrated coordinated lifting of a 60 mg and 13 cm long payload.These flight experiments illustrate better robot reliability compared to existing subgram aerial robots.More importantly, they represent exciting opportunities for future robotic studies of swarm behaviors in aerial insects.Traditionally, collective behaviors have been explored in terrestrial miniature robots [28] and mesoscale aerial robots. [29,30]However, most planning algorithms [31][32][33] for existing aerial robots [34,35] avoid mutual collisions because of the larger inertial effects.Owing to smaller size, our subgram MAVs are resilient against collisions, which open up new opportunities to investigate collision-friendly aerial swarm behaviors.In the past, biologists studied emergent swarm flight behaviors through observing aerial insects. [36,37]However, experiments with live insects have a large observational variance, and it remains difficult to isolate variables and test hypotheses under finely-controlled conditions.Developing at-scale aerial robot swarms will provide new tools for investigating these important biological questions.Our results not only address engineering challenges, but also open new opportunities for investigating fundamental scientific questions.Despite making progress in design, fabrication, and flight demonstrations, this work has several limitations.First, our flight controller does not consider robot interactions during flight and all flight trajectories are specified prior to the experiment.This lack of high-level real-time planning prevents the robot from demonstrating insect-like interactions, and it further limits the payload carrying capability.Future works can apply existing multiagent algorithms [38,39] in insect-scale aerial robots and study new behaviors such as mutual collisions and collective transport of objects in delicate environments.Second, due to hardware constraints, this work only flies two robots simultaneously, which remains limited compared to collective flight performed by insect swarms.Our experimental setup has a total of eight high voltage amplifiers, which is only sufficient for driving two robots that each requires four independent control signals.With more amplifiers, we expect our tracking system and flight controller will support simultaneous flight of up to ten robots.In the longer term, future studies can tackle the challenge of sensing and power autonomy through developing onboard sensors and power electronics.We envision future insect-scale MAVs will collectively perform insect-like tasks such as assisted pollination and hive construction.

Figure 1 .
Figure 1.Subgram aerial robots and flight demonstrations.A) An image that compares the prior (left) and the new (right) robots.The prior robot uses manually assembled carbon fiber airframe while the new robot uses stereolithographic 3D printing to improve fabrication scalability.B) A flight demonstration where the two robots in (A) collectively lift a 60 mg payload.

Figure 2 .
Figure 2. Robot design and 3D-printed components.A) An image of a 160 mg robot module.B) A CAD model of the robot module.The exploded view shows the robot module consists of an actuator, a 3D-printed airframe, 3D-printed connectors, transmissions, wing hinges, and wings.C) In a singlebatch fabrication, the Form 3 printer can make enough components for assembling 32 robot modules.D-F) Images of the 3D-printed items (left) and the autogenerated supporting bases for 3D printing (right).The static components include transmission connectors (D), single-unit airframes (E), and interunit airframes (F).The scale bars in (D-F) represent 1 cm.

Figure 3 .
Figure 3. Automated design and scalable fabrication of wing hinge and transmission.A,B) A design file and the corresponding lamination that contains 60 wing hinges.C) Top and side view photographs of the wing hinge.h w and h l are the user-defined design parameters.D) Cut files of nine material layers and the final release.KA, AD, and CF are abbreviations of Kapton, adhesive, and carbon fiber, respectively.E,F) A design file and the corresponding lamination that contain 48 robot transmissions.G) Top and side view photographs of the transmission.t h and t l are the design parameters.H) Cut files of 13 material layers and the final release.

Figure 4 .
Figure 4. Scalable fabrication of the robot wing.A,B) A design file and the corresponding lamination that contain 24 wings.C) Top view photograph of a 0.5 mg, 1 cm wing made of carbon fiber, adhesive, and polyester.D) Laser cut files of the carbon fiber layer, the adhesive layer, and the final release.E) A photograph of two wings that are made based on the prior laser cut files.The two wings have different spar widths, which lead to a 50 Hz mismatch of flapping resonance frequency.F) Comparison of prior[27] and new designs.The new cut file design substantially reduces realignment error and improves the consistency of wing inertia.

Figure 5 .
Figure 5. Experimental setup and robot characterization.A) The flight arena consists of control computers, a Vicon motion capture system, eight high voltage amplifiers, a high-speed camera, and safety switch boxes.B) A composite image of the robot flapping-wing motion at 1600 V. C) Tracked wing stroke and pitch motion corresponding to (B).D) A composite image of liftoff experiment where the robot carries a 100 mg payload.E) Measured robot lift-to-weight ratio as a function of driving voltage.

Figure 6 .Figure 7 .
Figure 6.A 5 s hovering flight.A) A composite image sequence that shows a 5 s hovering flight under feedback control.The blue dots indicate the location of the position setpoint.B-D) Robot x and y positions (B), altitude (C), and attitude (D) measurements.The setpoint of lateral position and attitude are 0 cm and 0°, respectively.The desired altitude function is superimposed with the measured trajectory in (C).

Figure 8 .
Figure 8.A 5 s beam lifting demonstration.A) A composite image sequence that shows the two robots collectively lifting a 60 mg, 13 cm long payload.The red and blue dots indicate the position setpoints of the original robot and the new robot, respectively.B-G) Robot lateral position (B,E), altitude (C,F), and attitude (D,G) measurements.Similar to the previous flight demonstration (Figure 7), the robot hovering setpoints are separated by 13 cm.

Table 1 .
Comparison of robot design and fabrication time.The reported time refers to the duration that requires active oversight from a researcher.This does not include the time related to automated processes such as 3D-printing or lamination.