Automated escape system: identifying prey's kinematic and behavioral features critical for predator evasion

Predation is a strong selective force that shapes various forms of defensive tactics of prey (Davies et al., 2012; Endler, 1991). One of the most common tactics when encountering a predator is the use of escape responses that produce rapid acceleration away from it (Cooper and Blumstein, 2015; Domenici and Hale, 2019). Numerous studies have explored the environmental and internal factors that influence the behavioral and kinematic variables associated with the escape response, such as reaction distance, escape trajectory, response latency and speed (Bateman and Fleming, 2014; Cooper et al., 2003; Kawabata et al., 2023; Meager et al., 2006; Stewart et al., 2014). These studies assume that these variables determine the success or failure of predator evasion. However, direct evidence linking these variables to evasion outcomes is limited (Dangles et al., 2006; Kimura and Kawabata, 2018; Stewart et al., 2013; Walker et al., 2005), especially when prey species show little variation in the variable being investigated. Understanding the relationship between these variables and successful predator evasion is crucial not only for comprehending the determinants of successful predator evasion but also for shedding light on the evolution of specific traits and the dynamics of predator–prey relationships on a larger scale.

One promising approach to linking the prey's behavioral and kinematic variables with the outcome of predator–prey interactions is the use of a simulated prey system, which allows the manipulation of the prey's kinematic and behavioral features, exposing them to real predators. For instance, Shifferman and Eilam (2004) and Ilany and Eilam (2008) manipulated the escape direction and flight initiation distance of prey items (i.e. dead mouse or chick) by attaching a string to the prey and manually pulling it from a distance, subjecting them to barn owl Tyto alba attacks. Szopa-Comley and Ioannou (2022) developed a robotic prey system and manipulated the repeated escape trajectories against predatory fish, either fixed to a specific direction or randomized within a 270 deg range (45–315 deg away from the predator). However, the response latency and speed of simulated prey in these studies are not comparable to real prey, making it difficult to establish a direct link between many of the prey's variables and survival. Creating a system that moves comparably to real prey animals remains challenging, especially for invertebrate and lower vertebrate species that respond quickly to predators (e.g. in less than 100 ms) and escape rapidly (e.g. exceeding 1.0 m s⁻¹) (Bullock, 1984).

Here, we have developed an automated escape system comparable to real prey species. The system can respond to a predator in less than 100 ms and achieve speeds exceeding 1.0 m s⁻¹ during escape. It automatically detects an approaching predator and pulls the prey away from the predator once the predator reaches a predetermined threshold distance. The system is relatively low cost (approximately 2000 US dollars), customizable, and versatile for various applications, including measurement of predator performance and eliciting escape responses in prey. This paper provides a detailed description of (1) the technical aspects of the developed system, (2) a performance test conducted to evaluate the system's ability to exhibit fast and rapid responses while maintaining consistency across successive trials, and (3) a case study investigating the effects of escape speed and reaction distance on the success or failure of evading live predatory fish.

The animal care and experimental procedures were approved by the Animal Care and Use Committee of the Faculty of Fisheries (Permit No. NF-0058), Nagasaki University, in accordance with the Guidelines for Animal Experimentation of the Faculty of Fisheries and the Regulations of the Animal Care and Use Committee, Nagasaki University.

The system comprised a USB camera (DMK33UX287, The Imaging Source Co., Ltd., Bremen, Germany), a laptop (DAIV 5N 20075N-CML, Mouse Computer Co., Ltd, Tokyo, Japan; CPU: Intel^® Core™ i7-10875H at 2.30 GHz), a microcontroller board (UNO R3, Elegoo Inc., Shenzhen, China) with a motor driver shield (SU-1201, EK Japan Co., Ltd, Fukuoka, Japan), a DC motor (RE-260RA, Mabuchi motor Co., Ltd, Chiba, Japan) with a pulley (radius 25 mm), and a weight connected to the pulley by a string (Fig. 1). The experimental arena was surrounded by double circular plastic walls (outer radius 248 mm, height 160 mm; inner radius183 mm, height 130 mm) (Fig. 1; Fig. S1). These walls were attached using three sets of two (outer and inner) PVC pipes (outer PVC pipes: inner radius 15 mm, height 35 mm; inner PVC pipes: inner radius 8 mm, height 70 mm) (Fig. 1; Fig. S1). The inner PVC pipe was attached to the inner plastic wall and the outer PVC pipe, and the outer PVC pipe was attached to the inner PVC pipe and the outer plastic wall (Fig. 1; Fig. S1). The inner wall and PVC pipes were positioned 15 mm above the bottom of the tank to allow the weight to pass through (Fig. 1; Fig. S1). The end of the string, running through the outer PVC pipe and the hole in the outer wall, was connected to the motor with the pulley outside the outer wall (Fig. 1; Fig. S1), creating an open line setup. Pulleys (radius 5 mm) were attached to the lower ends of the outer PVC pipes to reduce friction when moving the string. The string consisted of a polyethylene line (length 2.2 m) and a transparent carbon nylon line (length 1.0 m). We used two different lines because a carbon nylon line is transparent but is easily twisted, whereas a polyethylene line is the opposite. We attached a weight at the other end of the string as well as the prey to prevent the prey from moving. We used a fishing barrel swivel (0.17 g) as the weight to prevent it from rolling as a result of a twisted string. The escape direction of prey could be adjusted by using a different PVC pipe to pass through a string toward the motor.

The step-by-step process of programs for detecting an approaching predator and triggering the escape movement of the prey is shown in Fig. S2 (see Movie 1 for an example of how the system works). The USB camera was installed above the experimental arena to monitor the real-time movement of the predator and display the image on the laptop. Two circles – the search circle (outer circle) and reaction circle (inner circle) – were generated using a program superimposed onto the image (Fig. 1B,C). Inside the search circle, the uppermost point of the predator was detected and tracked (Fig. 1B,C). When the uppermost point of the predator was detected within the reaction circle, the laptop computer sent a signal to the microcontroller board. The custom program for the predator monitoring system was written in Python (v.3.7.7) with the OpenCV library (Opencv-contrib v.3.4.10.37). The USB camera recorded images in monochrome, but the output image format was RGB. Therefore, the program converted the images to a monochrome format. Subsequently, these monochrome images were converted to binary images using a threshold of 65. This process transformed the predator into a black object in the image, enabling the computer to detect the predator's contour. Before detecting the predator's contour, the program manually cropped the binary image using a search circle to exclude objects outside the search circle, such as the plastic wall and the tank equipment, which were also converted to black objects. Finally, the program detected the contour of the black object (representing the predator), converted it into a blob, and tracked the blob's uppermost point. This mechanism allows the detection of predators with non-uniform colors or complex shapes. The position and the radii of search and reaction circles could be adjusted in the Python program. The temporal resolution (i.e. frame rate) of the USB camera was limited by the program and exhibited fluctuations with a mean±s.d. of 238±39 frames s⁻¹. The custom Python program is available from GitHub (https://github.com/YuukiKawabata-Lab/PreyEscapeSystem).

The movement of the prey was controlled by the motor and microcontroller board with a custom program written in Arduino IDE (v.1.8.19) through a connection with the string between the weight (the prey) and the pulley. When the microcontroller board received the predator detection signal from the laptop computer, it sent movement commands to the motor, which rolled up the string and dragged the weight (with prey) (Fig. 1). The Arduino program provided the flexibility to adjust the timing (latency) and duration of the motor's rotation with millisecond resolution. The rotation speed of the motor could be regulated by Pulse Width Modulation (PWM) signal values ranging from 0 (minimum) to 255 (maximum). The custom Arduino program is available from GitHub (https://github.com/YuukiKawabata-Lab/PreyEscapeSystem).

To evaluate the performance and consistency of the developed system, we conducted 30 trials and analyzed the response latency and moving speed of the weight. The system was placed inside a rectangular acrylic tank (width 60 cm, height 60 cm, length 150 cm). Throughout the experiment, the water depth in the tank was maintained at 100 mm. The radii of the search and reaction circles were set to 153 mm and 40 mm, respectively, with the weight positioned at the center of these circles. Three different speeds (high, middle and low) were utilized in the motor control program, corresponding to PWM signal values of 245 (high), 40 (middle) and 11 (low). We conducted 10 trials at each programmed speed. In every trial, the motor rotation was programmed to displace the weight approximately 232 mm from its starting position towards the PVC pipe on the right side. The displacement of 232 mm was achieved by varying the running duration of the motor via trial and error for each speed. Delay time was not included in the program as our objective was to estimate the baseline response latency of the system. The system was activated by manually moving a dummy predator into the inner circle. The movements of the weight were recorded from above using a high-speed video camera (RX10IV, Sony Corp., Tokyo, Japan) with a frame rate of 480 frames s⁻¹.

The recorded videos were analyzed using Kinovea v.0.8.27 (www.kinovea.org). The response latency was determined as the duration between the entry of the dummy predator into the reaction circle and the initiation of weight movement. To determine the moving speed, the coordinates of the weight were digitized semi-automatically using Kinovea in each frame from its starting position to the inner plastic wall. Subsequently, we calculated the time-series speed of the weight from its positional data using the Lancoz smoothing method with custom R code.

To demonstrate the utility of the system, we exposed the controlled prey to a live individual predatory fish Coreoperca kawamebari Temminck & Schlegel 1843 (total length: 69.2 mm). We manipulated the escape speed (i.e. mean speed from the starting position to the inner plastic wall; Fig. 1) and reaction distance (i.e. radius of the reaction circle; Fig. 1) to investigate whether these variables influence the predation probability. This experiment was conducted using the same tank and system as in the performance test. The fish had been kept in a compartment (width 20 cm, height 25 cm, length 20 cm) made with transparent PVC punching plates. Live shrimp Neocaridina denticulata (De Haan 1844) (mean±s.d. total length 14.4±1.4 mm) attached to the weight with adhesive (Aron Alpha A Sankyo^®, Daiichi-Sankyo Co., Ltd, Tokyo, Japan) was used as prey. Live prey was chosen because the preliminary experiment showed that the predator did not chase or strike towards dead prey or artificial pellets. The prey's heads were positioned to face the entrance of the experimental arena. At the start of each trial, the compartment was positioned in front of the arena entrance, laid horizontally to allow the fish to enter the arena. As in the performance test, the radius of the search circle was set to 153 mm, and the system was programmed to move the weight 232 mm from its starting position towards the PVC pipe on the right side. A total of 33 trials were conducted, with the prey being randomly assigned seven different speeds (voltage values of 10–16, corresponding to mean speeds of 0.25–0.40 m s⁻¹) and eight different reaction distances (16–58 mm) to elicit the predator's movement. Because the actual escape speed cannot be measured when the predator successfully captures the prey (capture success: 19 trials, escape success: 14 trials), we measured the speed of the actual movements 5 times for each programmed speed before the trial (without the predator), and its mean value was used for the analysis. The water depth was kept at 100 mm and temperature was maintained at a mean±s.d. of 22.9±0.2°C throughout the experiment. We chose this temperature because it is close to the highest suitable temperature for this species (range 10–23°C) (Froese and Pauly, 2010).

In three cases, the predator touched the prey, but the predator finally failed to capture the prey. Because this study focused on kinematic and behavioral variables rather than the other variables (e.g. size, spines), these cases were regarded as successful capture.

To investigate the potential effects of sound and vibration generated by the motor on predator behavior, we used the same system without any bait attached, and the predator did not respond to the movement of the weight and string. This result suggests that the effects of sound and vibration are negligible.

In the performance test, the effect of three different programmed speeds on response latency was examined using one-way analysis of variance (ANOVA). In order to assess the consistency of the measurements, we calculated the coefficients of variation (CV) for the response latency, as well as the mean and maximum speed corresponding to each programmed speed. Consistency of the parameters were classified based on a previous study (Aronhime et al., 2014), defining excellent when CV was <10%, good when CV was 10–20%, acceptable when CV was 20–30%, and poor when CV was >30%.

In the case study, the effects of prey escape speed and reaction distance on predation probability were evaluated using a logistic regression analysis. Success and failure of predation were designated as 1 and 0, respectively, and used as the objective variable. Escape speed and reaction distance were considered as the explanatory variables. The significance of the variables was assessed by removing them from the model, and comparing the change in deviance using the likelihood ratio test with a χ² distribution. All analyses were carried out using R (v.4.0.5).

There was no significant difference in the response latency among the three programmed speeds (Table S1; ANOVA, F_2,27=0.82, P=0.450). The mean (±s.d.) response latency was 72.0±7.4 ms, which is comparable to the response latencies of visually evoked escape responses in different fish species (Batty, 1989; Cade et al., 2020; Dunn et al., 2016; Paglianti and Domenici, 2006) (Table S1; Fig. 2). With a CV of the response latency of 10.2%, the consistency of the response latency in this system is considered good (almost excellent).

Time-series moving speeds of the weight (prey) for three different programmed speeds are shown in Fig. 2F. The variations among trials were apparently small, and the CVs of the mean and maximum speeds for all three programmed speeds were less than 10% (Table S1). This suggests that the consistency of the moving speeds of the system is excellent. When the speed was set to low, the weight rapidly accelerated to about 0.3 m s⁻¹ in around 5 ms and maintained that speed. When the speed was set to middle or high, the weight reached the inner wall before its moving speed stabilized. For the middle speed set, the weight rapidly accelerated to approximately 0.7 m s⁻¹ in around 25 ms, followed by gradual acceleration with oscillations. For the high speed set, the weight rapidly accelerated to approximately 1.8 m s⁻¹ in about 15 ms, and then gradually accelerated to about 2.5 m s⁻¹ by the 40 ms mark. The speed then decelerated to 1.5 m s⁻¹ and eventually reached the inner plastic wall.

The oscillations observed at middle and high speed can be attributed to the extension and contraction of the string. According to the principles of classical mechanics, a force exceeding the static friction force is required to initiate weight movement. Therefore, the string attached to the weight would have been slightly extended before the onset of weight movement, which caused rapid acceleration after the onset of movement. The rapid acceleration would then contract the string, resulting in deceleration of the weight movement. This process is likely to have produced the oscillations in the speed of prey movement. Using an inextensible string and/or a weight made with a material that has less friction could help alleviate the oscillations in prey speed over time. Nonetheless, it is worth noting that these kinds of oscillation pattern are often observed in actual animal movements (Alexander, 2003; Voesenek et al., 2019).

The odds ratio, based on 1 s.d., was 0.64 for escape speed and 0.33 for reaction distance. These findings are consistent with previous studies that used live prey and predators (Kimura and Kawabata, 2018; Langerhans, 2009; Stewart et al., 2013; Walker et al., 2005).

Previous studies have used different simulated prey systems that manipulate the kinematic and/or behavioral variables of prey to study their interactions with live predators. The simplest method involves attaching a string to the prey and manually pulling it from a distance (Ilany and Eilam, 2008; Shifferman and Eilam, 2004). While this method is easy to implement, it is limited by the response latency of human visual processing, which is approximately 175–202 ms (Solanki et al., 2012). Additionally, the escape speed of the prey is influenced by human performance, which can lead to inconsistencies across successive trials. A virtual prey system, in which prey objects are projected onto one side of an experimental tank, has been used to examine the schooling parameters of prey by manipulating their movement trajectories (Duffield and Ioannou, 2017; Ioannou et al., 2012, 2019). However, live predators are generally attracted to the movement of virtual prey, and virtual prey are not programmed to respond to approaching predators. Therefore, this method is not suitable for situations where the prey remains stationary at first and rapidly initiates escape motion. Additionally, because predators are not rewarded for successful attacks on virtual prey, this system cannot address questions that require successive trials for individual predators to learn (e.g. random versus specific escape trajectories). A recently developed robotic prey system (Swain et al., 2012; Szopa-Comley and Ioannou, 2022), which responds to approaching predators by turning and escaping, can be successfully used to answer such questions. However, it has a longer response latency (approximately 700 ms), probably because of its complex system. In addition, the robotic system should be slower than our system as a result of the robotic prey's heavier mass. Consequently, it may be challenging for the robotic prey system to respond quickly and escape rapidly to evade actual predators, making it difficult to directly link the prey's variables to successful predator evasion. In the study by Szopa-Comley and Ioannou (2022), indirect metrics (i.e. time required for the predator to capture the prey) was used to examine the benefit of random escape directions over specific directions. In contrast, our system can control the prey's escape speed over a relatively wide range, including speeds that enable prey to evade actual predator attacks. Additionally, the prey reacts to approaching predators as quickly as real prey species (i.e. <100 ms). Therefore, our system is more suitable for research questions that need to elicit a predator's maximum performance, such as the establishment of a direct link between prey variables and the outcome of predator–prey interactions. While our current system is not designed for multiple individuals, the number of prey and escape direction choices can be easily increased by using multiple strings and motors, enabling simulation of grouped prey escaping in various directions.

One potential drawback of the developed system is that the prey's escape direction may be predictable to predators because of the presence of the string. This becomes especially problematic when the research question involves examining escape directions across successive trials (e.g. random versus specific escape directions) and predators can perceive or sense the string. To address this issue, one possible solution is to incorporate a magnet system similar to the one used in the robotic prey system (Swain et al., 2012; Szopa-Comley and Ioannou, 2022). In this setup, one magnet connected to a motor is positioned underneath the floor of the experimental arena, while another magnet is attached to the prey on the floor. However, it is important to note that using magnets may increase friction, resulting in slower prey movement and larger oscillation amplitudes in speed. Additionally, working with magnets could interfere with the magnetic detection fields of some fish and would not allow lure-type prey to float and move freely, as a result of this magnetic attraction. Another potential solution is to use multiple strings placed underneath the weight, which could prevent predators from predicting the escape direction. Considering the above, the solution will depend on the biology and characteristics of the subject of study. Further research is necessary to investigate whether predators can perceive transparent strings and to evaluate the effectiveness of modified systems, such as those utilizing magnets or multiple strings.

Our system has potential to be applied for a range of objectives. First, the response latency of our system is adjustable via the Arduino program, and the escape trajectory can be altered by adding more PVC pipes and threading the string through them. This allows investigation of how variation in these parameters affects the outcome of predator–prey interactions. Second, as implied by the case study, our system is useful for measuring behavioral traits of predators such as reaction time, approach speed, interception strategy, learning and plasticity. The third potential application is the elicitation of escape responses in prey species. Many studies have employed artificial stimuli, such as dummy predators, air-puffs or dropping balls, to trigger escape responses in prey (Camhi and Tom, 1978; Kawabata et al., 2023; Marras and Domenici, 2013; Meager et al., 2006). In these studies, observers typically wait for the prey to approach a specific location or force the prey to stay at that location before triggering the artificial stimulus. Our system can be modified to efficiently collect data in such cases by placing the reaction circle in front of the artificial stimulus and connecting the microcontroller board to the device that generates the artificial stimulus. These examples demonstrate the versatility of our device, which can be utilized for investigating various questions in animal behavior.

In this study, we have introduced a novel prey escape system that demonstrates fast and rapid responses to predators while maintaining consistency across successive trials. Our system allows manipulation of various kinematic and behavioral variables of prey, enabling us to establish a direct link between these variables and successful predator evasion. Given the design and functionality of our system, it would also be applicable to studies involving terrestrial predators. In conclusion, our prey escape system holds great potential for advancing knowledge of predator–prey interactions in various ecological contexts.

We sincerely thank the staff and the students in the Kawabata Laboratory for their daily support in our research activities. We appreciate the valuable comments from two anonymous reviewers, which significantly improved the manuscript.