Wind Source Localization System Based on a Palm-Sized Quadcopter

Abstract

In this study, we implemented a compact wind direction sensor on a palm-sized quadcopter to achieve wind source localization (WSL). We designed an anemotaxis algorithm based on the sensor data and experimentally validated its efficacy. Anemotaxis refers to the strategy of moving upwind based on information on the wind direction, which is essential for tracing odors propagating through the air. Despite the limited research on quadcopter systems achieving WSL directly through environmental wind measurement sensors, debate remains regarding the relationship between sensor placement and the anemotaxis algorithm. Therefore, we experimentally investigated the placement of a wind direction sensor capable of estimating wind source direction even when propellers are rotating. Our findings demonstrated that placing the sensor 50 mm away from the enclosure of the quadcopter allowed accurate wind direction measurement without being affected by wake disturbances. Additionally, we constructed an anemotaxis algorithm based on wind direction and speed data, which we integrated into the quadcopter system. We confirmed the ability of the quadcopter to execute anemotaxis behavior and achieve WSL irrespective of environmental wind strength through wind source localization experiments.

1. Introduction

This paper proposes a system for wind source localization using a palm-sized quadcopter.

Estimating the direction of the surrounding water or air flows and moving accordingly constitutes a crucial survival strategy [1,2] for aquatic organisms (rheotaxis) [3] and flying animals (anemotaxis) [4]. These strategies are essential for tracking chemicals that are primarily dispersed through water or air, whose uneven and unpredictable distribution makes them challenging to localize using only dynamic chemical information. Thus, organisms enhance their tracking capabilities by employing strategies that oppose water or air flows.

Identifying chemical sources supports critical tasks such as detecting gas leaks primitively, locating hidden explosives, and searching for victims in disaster zones. Consequently, research into autonomous robot-based chemical-source exploration has been actively promoted [5,6]. Therefore, it is expected that the task of odor source localization be performed indoors, such as in airports or collapsed buildings, where the global navigation satellite system (GNSS) is difficult to use. In the case of autonomous systems, most UAVs (Unmanned Aerial Vehicles) and robots use GNSS to obtain accurate location information for navigation. Navigating autonomous robots in situations where GNSS is not available is a problem to be solved, and one method is to employ probabilistic statistical methods to estimate position using only local information. One of the typical methods considered is SLAM (Simultaneous Localization and Mapping) [7], and recently, it has been proposed to utilize a coalition formation game model [8] or multiobjective motion planning strategy [9] to navigate autonomous robots, including UAVs, to a certain waypoint. Moreover, methods that use LiDAR depth information to understand spatial conditions and improve positioning accuracy have been proposed in recent years [10]. These methods are suitable for guiding autonomous robots to general waypoints, but for the problem of locating a particular source (e.g., odor, wind), the utilization of local sensory information can provide efficient navigation in situations where GNSS is not available. This method is represented as a bio-inspired method. Although animals do not have global coordinates such as GNSS, they can achieve efficient navigation by effectively utilizing multiple sensory information that can be acquired from local space. In fact, it has been reported that integrating multiple sensory inputs to determine behavior, as animals do, improves search efficiency in locating an odor source [11,12,13]. To employ this bio-inspired method, we need to implement sensors to correctly acquire local information. In the context of rheotaxis or anemotaxis, it is important to correctly implement sensors to detect water streams or air flows.

Furthermore, research is being conducted on implementing rheotaxis and anemotaxis in robots [14,15] to support the chemical source detection, which can be achieved with underwater and ground-running robots. In these studies, sensors that can directly detect water or wind flow are implemented on the robot, and this is achieved using relatively simple control laws. However, anemotaxis has not yet been achieved in a quadcopter. A quadcopter generates a strong airflow in order to fly. This airflow is referred to using the word “wake” [16], which has also been used in this paper. To achieve anemotaxis with a quadcopter, distinguishing between the environment wind and the wake and determination of the appropriate direction of movement are crucial. A method has been proposed to measure the wind environment with the quadcopter to estimate wind direction from the relationship between the attitude of the quadcopter as it moves and its attitude that changes due to being blown around by the wind [17,18]. This method can be used to estimate wind direction using an actual quadcopter. However, a problem remains with the estimation accuracy decreasing in environments where the wind direction changes rapidly. To address this issue, recent studies have reported the successful attachment of a wind sensor to a medium-sized quadcopter for observing natural wind information [19,20,21]. This study minimizes the impact of the wake of a quadcopter by positioning the wind sensor at a height away from the central axis of the quadcopter. However, because the primary objective was to observe natural winds, the quadcopter was manually maneuvered to observation locations, measuring environmental winds during hovering without autonomous functionality or anemotactic capability. Consequently, research on the implementation of the anemotaxis ability in the quadcopter is still controversial and underdeveloped.

Therefore, this study aims to implement anemotaxis in a quadcopter, enabling autonomous wind source exploration. By achieving the above objectives, this paper demonstrates that (1) by attaching a wind direction sensor to the palm-sized quadcopter, it is possible to distinguish between wake and environmental wind, and (2) the important biological localization strategy of anemotaxis can be reproduced by the quadcopter, thereby contributing to having more options to realize localization strategies for UAVs. In particular, we propose integrating lightweight wind direction sensors on a palm-sized quadcopter and developing an algorithm to navigate toward wind sources based on wind direction and speed information. The experiments were conducted under varying environmental wind speeds, evaluating a success rate of source localization and upwind movement efficiency to validate the wind source localization system.

2. Problem Statement

This study aims to implement anemotaxis capabilities in a palm-sized quadcopter and achieve wind source localization (WSL). As quadcopter size increases, benefits include longer flight times and increased payload capacity; however, larger quadcopters face challenges such as difficulty navigating indoor spaces [22] and potential harm to individuals. Therefore, this research adopts the palm-sized quadcopter platform (Tello, Ryze Technology, Shenzhen, China).

To measure environmental winds using quadcopters, environmental winds must be distinguished from wakes generated during flight. Two main methods are utilized for quadcopter-based wind observations: The first is to estimate wind indirectly from the relationship between the ideal attitude—which changes according to the control input of the quadcopter—and the actual attitude. The second is to measure wind directly by attaching the wind sensor in a location where the influence of the wake is minimal. Indirect estimation methods avoid additional wind sensors, thus benefiting payload capacity, although they may fail to accurately estimate wind speed and direction under unexpected conditions. In contrast, direct measurement using wind sensors ensures the accurate acquisition of wind speed and direction as long as the sensors remain operational. This study adopts the direct measurement method using a wind direction sensor to achieve WSL under varying wind conditions.

Past research implementing wind sensors on a quadcopter primarily proposed them as tools for monitoring environmental winds capable of measuring one-dimensional wind speeds. A human pilot was in control of navigating to specific destinations because its research goal was monitoring the environment. This study aims to detect wind speed and direction using a sensor, developing a system that autonomously makes appropriate decisions based on this information. To achieve this objective, the following components are integrated into the system:

• The arrangement of a wind direction sensor to minimize wake interference and detect wind direction effectively.
• The development of an anemotaxis algorithm capable of making the appropriate decisions based on wind speed and direction information.

Regarding sensor arrangement, past research suggests positioning them centrally and at a vertical distance from the quadcopter [19,20]. However, the exact height adjustment for different quadcopter sizes remains unknown, and its position should be determined experimentally. Although the sensor directly measures wind speed and direction, considering the possibility that the effects of wakes cannot be eliminated, we must develop an anemotaxis algorithm that prevents movement in the wrong direction by utilizing wind speed information comprehensively in addition to wind direction. The combination of the proposed sensor arrangement and anemotaxis algorithm is experimentally evaluated by conducting WSL experiments in environments with different wind speeds. Assuming that localization is not achieved in a windless environment and the localization success rate is 70% or higher in wind environments with different wind speeds, the system is deemed to have been successfully constructed. This 70% standard was set as the same localization success rate as in the initial research into odor source localization using a palm-sized quadcopter [22].

3. Development of a Wind Source Localization Quadcopter System

3.1. System Configuration

Figure 1 illustrates the exterior and schematic diagram of the wind source localization (WSL) quadcopter system constructed in this study. The quadcopter utilizes the palm-sized quadcopter, Tello. Tello has an internal controller for stable flight, which this research utilizes for basic flight control. The quadcopter system includes a microcontroller for quadcopter control and a wind direction sensor. The system employs the compact and lightweight TinyPICO (Unexpected Maker, Melbourne, VIC, Australia) for the microcontroller. Although TinyPICO is a small microcontroller, it has sufficient specifications to process wind sensor values and control the quadcopter (Figure 1c). Although TinyPICO features Wi-Fi and Bluetooth, it has a single antenna, making simultaneous Wi-Fi and Bluetooth use impossible. Consequently, two TinyPICO boards have been implemented: one for quadcopter control via Wi-Fi communication and sending velocity commands, and another for data logging.

The wind direction sensor adopts a lightweight 1.2 g hot-wire anemometer (HWD-18V-ONE, Hort-Plan, Osaka, Japan). At the tip of the wind direction probe shown in Figure 1, multiple heaters and temperature sensors are integrated to calculate wind direction based on their readings. Wind direction and speed data are transmitted via serial communication at one-second intervals to the TinyPICO. Wind direction data are read by the first TinyPICO and transmitted to the microcontroller (second TinyPICO) for quadcopter control. An additional battery powers both microcontrollers and the wind direction speed sensor. The weight of the control circuit, including the additional battery, is 24.4 g, which is well below the payload that Tello can fly (<50 g) [23].

3.2. Arrangement of Wind Direction Sensor and Performance Verification Experiment

The wind direction sensor is placed at the center of the quadcopter. However, the height at which the wind direction sensor should be installed is unknown; therefore, the height at which the wind direction sensor should be installed must be determined. Hence, as shown in Figure 2a, we verified the mounting of the wind direction sensor at distances of 50 and 100 mm from the quadcopter propellers to assess its ability to measure environmental winds. The experiment was conducted with the quadcopter fixed to a monopod, and measurements were taken 3 m from the wind source. The wind speed at the quadcopter setup location was measured using a commercially available anemometer (testo 405i, Testo SE & Co. KGaA, Titisee-Neustadt, Germany), which was calibrated beforehand and was low at approximately 0.1 m/s.

Wind direction was measured for 60 s at each sensor height during the hovering state with rotating propellers, and the results of the average wind direction and standard deviation are illustrated in Figure 2b. The vertical and horizontal axes of Figure 2b represent measured wind directions and true wind directions, respectively. The black circles in Figure 2b represent ideal values, while blue and red colors denote conditions for lower (50 mm) and higher (100 mm) sensor heights, respectively. Figure 2b indicates that the sensor height of 50 mm results in less error from the true wind direction across all angular conditions.

Based on these findings, the wind direction sensor was installed at a height of 50 mm to detect environmental wind direction information in the proposed system. To further verify the detection of more detailed wind direction information with this sensor arrangement, measurements were conducted in increments of 22.5° from 0 to 360°. Additionally, experiments were conducted under conditions of propeller rotation switched on and off to assess the extent of wake influence, maintaining consistent wind source distance and speed conditions as in Figure 2c. Figure 2c illustrates the average wind direction and standard deviation measured over 60 s at each wind direction, with the vertical and horizontal axes representing measured and true wind directions, respectively. The black lines in Figure 2c denote ideal conditions, while blue and red data points represent experiment results with propellers on and off, respectively. The results indicate that the system can detect wind source directions accurately with minimal influence from propeller rotation, even with fine variations in wind source orientation.

3.3. Design of the Anemotaxis Algorithm

We propose an anemotaxis algorithm that localizes to the wind source based on wind direction and speed data obtained from the wind direction sensor. Figure 3a shows the flowchart of the proposed algorithm. The algorithm is accomplished by repeating the following steps: (1) observation of the environmental wind, (2) determination of the direction of movement, and (3) movement. During the environmental wind observation phase, the quadcopter measures the environmental wind while hovering because the wind is generated as the quadcopter moves. Since the sampling rate of the wind sensor is 1 s, the environmental wind is observed by hovering for 10 s. After observing the wind direction information, the mode value is calculated to determine the wind source direction. If the wind speed is not fast, no environmental wind is likely to occur, and therefore, it is not included in the mode value calculation. In this study, we set the wind speed threshold to 0.27 m/s empirically. If the number of data points not included in the mode calculation increases, the quadcopter selects the random search state because it is close to a windless environment. After the direction of the natural wind is estimated by the mode value calculation, the direction of movement is determined. The four directions of movement are the front, back, left, and right. Therefore, the obtained wind direction information was divided into four parts, as shown in Figure 3b, to determine the direction of quadcopter movement. The quadcopter moves 0.2 m per movement, and wind source localization is achieved by repeating this process.

4. Wind Source Localization Experiment

4.1. Experimental Condition

The WSL experiments were conducted in an indoor area depicted in Figure 4a, measuring 4.0 m × 3.0 m × 3.0 m. The experimental area was equipped with a ceiling camera (BSW505MBK, Buffalo Inc., Aichi, Japan, 30 fps) capable of observing the entire area and recording the trajectory of the quadcopter. The recorded quadcopter videos were analyzed using a deep-learning-based video analysis tool [24].

Because this study is the first step in WSL, the problem was simplified and limited to a search in a two-dimensional plane. Therefore, the altitudes of the quadcopter and the wind source were set at the same height. The altitude was approximately 1 m above the ground, and the quadcopter began its search 2 m away from the wind source.

A fan (DR-HTF007, Dreo, Hong Kong, China) was installed as the wind source, configured with the following three scenarios:

• Scenario 1: Windless condition (fan powered off);
• Scenario 2: Wind speed under 2 m/s;
• Scenario 3: Wind speed over 2 m/s.

The wind speeds for Scenarios 2 and 3 are the set values for the wind source. To figure out the environmental wind profiles for Scenarios 2 and 3, wind speed maps were constructed according to the procedure shown in Figure 4b. The wind speed distribution for each scenario was measured in the x and y directions at 0.5 m intervals using a commercially available wind sensor (testo 405i, Testo SE & Co. KGaA, Titisee-Neustadt, Germany) and calculated by linear interpolation. Wind speed was measured at each measurement point for 60 s, and the mean value was defined as the wind speed at that point. Environment wind maps at the altitude where the quadcopter conducted localization for Scenarios 2 and 3 are shown in Figure 4c,d. Figure 4c,d clearly show that wind speeds decrease with distance from the wind source in both scenarios. As shown in Figure 4c,d, the wind speed decreases as one moves away from the wind source in both scenarios. The wind speed at 2 m, which is the starting point of the search, is approximately 0.5 m/s, which is relatively low in both scenarios. Ten repeated experiments were conducted under these conditions. Successful localization was defined as the quadcopter moving to a position sufficiently close to the wind source. Localization failure was defined if more than 2 min elapsed from the start of localization or if the quadcopter moved outside the exploration area. The quadcopter system is evaluated based on metrics including localization success rate, localization time, and upwind movement rate derived from the experimental results.

4.2. Experimental Results

The trajectories of each scenario are depicted in Figure 5a–c. Figure 5a–c correspond to Scenarios 1–3, respectively (see Supplementary videos). We indicate the start and end positions of WSL using “■” and “•”, respectively. The red areas in Figure 5a–c represent the goal areas, while the blue and black lines denote successful and failed localizations, respectively. The trajectory shows that in Scenario 1, the quadcopter moved to various places and exceeded the time limit in most cases. In Scenarios 2 and 3, while instances of incorrect movements occurred, the quadcopter was confirmed to move upwind direction in most trials.

To statistically evaluate these trajectories, the localization success rates, localization times, and upwind movement selection rates (hereafter referred to as upwind rates) for 10 repeated experiments in each scenario are shown in respective Figure 6a–c. The localization success rate was 0% in Scenario 1. In Scenarios 2 and 3 with environmental winds, success rates were 70% and 90%, respectively, indicating a higher success rate under the stronger environmental wind conditions of Scenario 2. Significant differences between Scenarios 1, 2, and 3 were observed in localization success rates (Fisher’s exact test, p < 0.05). The calculations of localization times for successful localizations in Scenarios 2 and 3 showed no significant difference despite variations in environmental wind strength. Thus, we found no significant increase in localization time due to reduced movement speeds influenced by environmental winds (Wilcoxon rank-sum test, p < 0.05). The degree to which the quadcopter selects to move upwind during the localization behavior is evaluated by comparing the upwind rate. The upwind rate is calculated as follows:

$$\begin{array}{ccc}\hfill \mathrm{Upwind}\mathrm{Rate}& =& {\displaystyle \frac{\mathrm{Upwind}\mathrm{movement}}{\mathrm{All}\mathrm{movement}}}.\hfill \end{array}$$

(1)

The calculation of the upwind rate based on this equation showed a significant difference between environments with no wind and those with environmental wind. However, no significant difference was observed between Scenarios 2 and 3, confirming no significant change in behavior selection with the strength of the environmental wind (Steel–Dwass test, p < 0.05).

Overall, these results confirm that even under the low-speed conditions of Scenario 2, the quadcopter could extract only the ambient wind from the wake and move in the upwind direction. The trajectory in the windless condition of Scenario 1 shows that the quadcopter always moves randomly because it cannot detect the ambient wind, and the search fails because the quadcopter moves out of the search area. This indicates that the combination of the wind direction sensor arrangement and the anemotaxis algorithm can be used to detect the ambient wind and localize the wind source.

Next, we analyzed the wind direction sensor values during WSL. Figure 7a shows histograms of wind speed values measured in 10 repeated experiments for each scenario. The blue lines in Figure 7a represent the thresholds for calculating the mode of wind speeds. The different colors in the histograms represent each scenario, illustrating varying distributions of measured wind speeds under different wind conditions.

Even in the windless condition (Scenario 1), slight wind speeds were measured due to wake effects, suggesting that threshold processing suppressed anemotaxis behavior, resulting in random behaviors. Additionally, rose plots of measured wind direction values are shown in Figure 7b–d. The amplitude of the rose plots represents occurrence frequencies, with red indicating the mean value. The 0-degree direction represents the upwind direction. Although a bias exists toward the rear direction in the windless condition, almost every azimuth has a value. Conversely, more wind directions were detected around the 0-degree direction under windy conditions. Moreover, in Scenario 3, with higher wind speeds, the rose plot shows a sharper distribution toward the upwind direction.

These findings confirm that even with a palm-sized quadcopter, minimizing the wake effects allows for the accurate detection of wind direction and speed. It also underscores the importance of combining wind speed information to determine movement direction, particularly in windless environments with a higher risk of moving in incorrect directions due to wake effects.

5. Conclusions

In this study, we constructed a system using a palm-sized quadcopter equipped with an anemotaxis algorithm to achieve WSL. Many organisms elicit the ability to move against flows, a crucial strategy for locating food sources or mates. This biological strategy has been successfully implemented in terrestrial and underwater robots to enhance odor source localization performance. However, the detection of environmental winds was believed to be challenging using quadcopters capable of three-dimensional flight due to the influence of wake effects. Recent research advances have shown that installing anemometers and wind vanes away from the quadcopter body allows for the effective monitoring of environmental winds. Thus, in this study, we integrated a wind direction sensor capable of observing environmental winds on the palm-sized quadcopter and implemented an anemotaxis algorithm to construct a system capable of localizing a wind source.

We installed a lightweight heat-probe-type wind direction sensor weighing approximately 1.2 g on the quadcopter, with its position and height determined experimentally. Based on insights from prior studies, we placed the wind direction sensor at the center of the quadcopter but adjusted its height according to the quadcopter size to account for variations in wake strength. We attached the sensor at 50 and 100 mm heights in experiments to investigate whether the true wind source direction could be accurately estimated under conditions where the propeller rotation generated the wake. The results demonstrate that even at a lower sensor height (50 mm), the accurate estimation of the true wind source direction was feasible, thereby determining the appropriate sensor arrangement to achieve anemotaxis.

We developed an anemotaxis algorithm based on wind direction and speed data from the wind direction sensor. Due to the inherent characteristics of the wind direction sensor, which continues to output values even in the absence of environmental winds, we calculated the mode value from wind direction data when a consistent wind speed was observed to determine the direction of movement. This approach effectively suppressed erroneous movements due to false detections and enabled accurate localization toward the direction of the wind source.

Through WSL experiments conducted in three environmental scenarios, we confirmed the capability of the system to localize the wind source independently of wind strength. Additionally, in scenarios without wind, the constructed anemotaxis algorithm operated by entering an exploration mode involving random movement due to the lack of correct wind direction information.

Thus, by implementing the wind direction sensor in areas of minimal wake influence on the palm-sized quadcopter, we successfully detected environmental winds and constructed a system capable of WSL.

This technique is a navigation strategy that works in windy environments and may not work in enclosed, windless environments. To develop a practical system, it will be important to switch navigation strategies depending on the environmental conditions, such as integrating other sensory information, utilizing GNSS information, and using statistical position estimation methods. In addition, because the focus of this study was to clarify how reliable the information from the wind direction sensor is, we employed the simplest behavior decision algorithm that directly uses the information from the sensor. However, in environments where wind direction constantly varies or multiple wind sources exist, this simple algorithm is insufficient. Hence, in the future, we will develop an algorithm that utilizes the history of wind speed or direction data to determine the direction in which wind information is likely to be obtained from an information-theoretic perspective.