Motivation

The main motivation behind the development of the box was to introduce a second layer of specificity to our system for targeting invasive species. Our initial off-target analysis indicated that the shRNA designed to silence essential genes in the Asian hornet (Vespa velutina) would also potentially affect other related species, including the European hornet (Vespa crabro). Since the European hornet is a native species across Europe and plays a critical role in pollination and pest control (Pusceddu et al. 2022), our box was a vital part to avoid harming local ecosystems.

Given the European hornet's ecological importance, we needed to create a mechanism that would ensure that only the invasive Asian hornet could access the bait containing the shRNA. To that end, our box was designed to restrict entry to non-target species through an image recognition software. By designing a species-specific system, we aim to selectively reduce the population of the Asian hornet without harming native species such as the European hornet. This added layer of precaution minimises ecological risks and enhances the safety of our engineered bacterium.

Figure 1: The two layers of specificity to avoid targeting native species.

The first layer represents our shRNA, which will prevent most other species from being targeted. The second layer represents the image recognition software, which will prevent European hornets from being targeted.

To achieve the selectivity with our hardware, we needed to implement the following components: a reliable recognition mechanism, a physical separation from the contaminated bait, and a mechanism to open and close the box. Designing this system required at least one motion sensor, a camera for identifying the Asian hornet, a rigid stationary structure (box), movable parts (doors), and servo motors to enable the movement.

Design process

When initiating the design process for the detection box, we first attempted to adapt the design created by VespAI authors (O’Shea-Wheller et al. 2024). Their design included a dome where a camera was positioned. This dome, made from a white translucent material, was intended to provide uniform lighting on the platform, enabling the image recognition software to more accurately detect the Asian hornet by reducing shadow variability, and consequently, minimizing noise in the data. We modified this design by adding a closed compartment beneath the dome to contain our protein bait (Figure 2).

We presented our design to Federico Cappa, a postdoctoral researcher with expertise in Asian hornet behavior. He noted that our initial design was not specific enough to the Asian hornet, as it would still allow the European hornet to access the bait if both species were present in the dome simultaneously. This feedback indicated that further refinement was needed to improve the species specificity of the system.

Figure 2: Initial Designs Inspired by VespAI

Both designs feature a dome that provides refracted uniform lighting for optimal image recognition, supported by four structural pillars. The key difference lies in the bait access mechanism: Design 1 incorporates a sliding door, while Design 2 uses a trap door system.

After consulting with Federico Cappa, we revised our design to resemble a tube, which separates the initial image recognition compartment from the second compartment containing the bait (Figure 3). He provided several key suggestions to improve the design and ensure that Asian hornets are attracted to the box and feed on the protein bait. Firstly, he recommended widening the entrance to prevent hornets from feeling intimidated or hesitant to enter. He also suggested enlarging the bait compartment to reduce stress on the hornets; a cramped space might cause them to leave without feeding. Finally, he advised constructing the exit door with transparent material to allow light through, making it easier for hornets to exit, as they are naturally attracted to light.

We retained several elements from the original VespAI design. These include the homogeneous lighting (diffused, non-direct light), which minimizes shadow and color variations to enhance the accuracy of the image recognition software. We also kept the raised middle section to ensure the camera captures the hornets from a similar distance to that used in training the AI model.

Figure 3: Improved design of the box

This revised design includes several enhancements to make the delivery of shRNA more efficient.

For the final version of the design, we carefully considered the materials for its construction. We opted to 3D print the design, and to simplify the printing process, we modified it from a rounded shape to a more angular form (Figure 4). This adjustment allowed us to avoid creating support structures during the 3D printing process, optimizing our resources.

Figure 4: Final design

The final design of our box functions as follows:

• An insect enters the box through the entrance, which remains open by default.
• The image recognition system identifies the insect using a camera. If it is not an Asian hornet, or if both an Asian hornet and a European hornet are present in the chamber, no mechanism is triggered.
• However, if the insect is an Asian hornet alone, it will trigger the entrance door to close and simultaneously open the door leading to the bait.
• The hornet takes a piece of bait to bring back to the nest.
• While flying up towards the light, the motion sensor detects the hornet, causing the exit door to open, allowing it to leave the box.
• The box resets to its default state, where the entrance door is open, and both the door to the bait and the exit doors are closed.

Details implemented in our final design:

• For printing materials, we chose PLA because it was more available during the prototype phase, but for future versions to be tested outdoors, we plan to use PETG, which is more UV resistant.
• The compartment containing the bait is a 15x15x10 cm cube, providing enough space for the hornets to avoid feeling stressed.
• The environment inside the box remains consistent, allowing the AI to perform well. We achieved this by incorporating homogenous lighting (diffused, non-direct light) to reduce variations in shadows and colors.
• The last door is made from transparent polystyrol to allow light to pass through, enabling the hornets to fly towards the motion sensor as they are naturally attracted to light.

UNILausanne VESPA TEAM: Protein Bait Container Integrated with Image Recognition Software

Technical details of the hardware

Component	Technical Specifications	Use
Raspberry Pi 5 / 8 GB	2.4GHz quad-core, 64-bit Arm Cortex-A76 CPU with 512KB L2 caches and a 2MB shared L3 cache | VideoCore VII GPU for enhanced graphics performance | Dual-band 802.11ac Wi-Fi and Bluetooth 5.0/BLE for robust wireless connectivity | Gigabit Ethernet with PoE+ support for high-speed wired connections	Development board responsible for receiving input from the Pi Camera and PIR sensor, and outputting signals to servos and motion detector.
Pi Camera Module 3	12MP IMX708 Quad Bayer sensor | Resolution: 11.9 megapixels | 1.4µm × 1.4µm pixel size | output: RAW10 | integrated IR cut filter | Phase Detection Autofocus	Pi compatible camera that takes a series of photos at regular intervals.
Joy-it Motion Sensor	SBC-PIR	Infrared based motion detector used for triggering the opening of the last door.
Purecrea 16-channel PWM / Servo driver I2C	Operating voltage Logic: 3.3 to 5V | 0x40 I2C standard address | Channels: 16 Channels | Resolution: 12-Bit | Maximum PWM frequency: 1.6KHz	Driver to ensure that the right amount of voltage and current arrive to the motors.
WaveShare Micro Servo 9G / SG90	Type: WS-SG90 | Torque: 2.0kg/cm(4.8V), 2.2kg/cm(6V) | Speed: 0.09s/60°(4.8V), 0.08s/60°(6V) | Dead band: 7us | Gear: POM | Operating voltage: 4.8V to 6V DC	Servo motors used to open the doors of the box.
Duracell Mignon (AA) battery MN1500 Plus	Battery type: Alkali | Battery Application: AA batteries | Voltage: 1.50 V	Power supply for the servos and motion sensor.
Goobay Battery holder	8x Mignon (AA) cable (L x W x H) 63 x 58 x 29.5 mm (81218)	Connects batteries with servo driver to provide power supply for servos and motion sensor.

When thinking about the construction and feasibility of our box, we had to carefully consider which motors, which motion sensors, and which Raspberry Pi to use. To control the Servo DC motor, we utilized the gpiozero package, a Python library designed specifically to interact with the Raspberry Pi’s general-purpose input/output (GPIO) pins (“2. Basic Recipes — Gpiozero 2.0.1 Documentation,” n.d.). This allowed us to effectively control the motors through Python programming.

Why Servo motors?
Initially, we considered using stepper motors due to their precision and the early designs that required exact rotor positioning. However, as we finalized our design, we realized we could opt for a simpler and more efficient solution. We chose small servo motors because they are versatile, cost-effective, and energy-efficient, which suited our project’s needs perfectly.

The decision to use a Raspberry Pi was straightforward. Although we briefly considered using an Arduino, we opted for the Raspberry Pi because of our greater familiarity with Python rather than C++, which is required for programming on the Arduino. The Raspberry Pi's ability to be easily programmed in Python made it the more logical choice for us.

In our initial designs, we did not include a motor driver since we had only one small motor that could be powered and controlled directly by the Raspberry Pi. However, as the project progressed and we incorporated a second motor, we decided to include the motor driver. The role of the driver is to ensure that the motors receive the correct voltage, which in our case was 5V, ensuring smooth and reliable operation.

Figure 5: Simplified Schematics.

Future improvements

Without field testing, it is challenging to identify all the areas that require adaptation for optimal functionality. However, several enhancements have already been considered to improve the design's efficiency, durability, and usability. To increase the autonomy of our box, integrating solar panels on the roof is a priority. This would enable the system to operate independently without frequent battery replacements or external power sources, making it more convenient for beekeepers and reducing the time they need to invest in maintenance.

Currently, the motors were selected primarily based on cost considerations. For future iterations, we would explore more efficient motors with lower power consumption and higher durability to enhance the box's performance and longevity. This would be especially important if the system is to be used in remote locations with minimal human intervention. Additionally, while the current prototype uses PLA, we recommend switching to PETG for outdoor use due to its UV resistance and durability. This would help the box withstand harsh weather conditions, reducing the risk of degradation over time. Additionally, other materials such as polycarbonate could be considered, further increasing the lifespan of the hardware.

Next, to ensure reliable operation in various environmental conditions, we would test the box under different temperature and humidity scenarios. This could lead to adjustments in the design, such as incorporating ventilation or insulation features, to maintain optimal internal conditions for the hardware and image recognition software.

As we would consider beekeepers to be our endpoint users, we thought of implementing a communication module for remote monitoring and updates, which could send alerts to the beekeeper if a malfunction occurs or when the bait needs to be changed. The addition of clear instructions and potentially a companion app could guide users through the setup and troubleshooting process, making the system more user-friendly. However, since cost-effectiveness remains a priority, all of these improvements require a balance between functionality and cost-efficiency.

These improvements would collectively enhance the system's reliability, usability, and effectiveness in the field, making it a more robust tool for targeting invasive species while preserving native ecosystems.

References

• O’Shea-Wheller, Thomas A., Andrew Corbett, Juliet L. Osborne, Mario Recker, and Peter J. Kennedy. 2024. “VespAI: A Deep Learning-Based System for the Detection of Invasive Hornets.” Communications Biology 7 (1): 1–11. https://doi.org/10.1038/s42003-024-05979-z.