To guide the design of the sensorized wall we conducted eight in-depth, semi-structured expert interviews (35). Six experts were identified from the climbing field: two climbers, two climbing instructors, and two physiotherapists. Two more experts were selected from a research institute (engineers/climbers). They were asked to share their personal context-specific knowledge instead of providing official association statements (36). The qualitative approach with semi-structured interviews provides open-ended answers, not quantitative data. These interviews, and the needs that we identified, were turned into design requirements (see Table 1) through a further round of literature review and interaction between the research team and the experts.

One of the common needs highlighted by the experts was to obtain a quantitative measure related to weight distribution, and more in general a measure directly or indirectly related to body pose, and athletic and motor performance during regular climbing tasks in a variety of climbing scenarios, with a variety of hold shapes and arrangements. This was however combined with the need to avoid wearable devices and modifications to the surface of the wall that could interfere with athletic gestures. To meet this requirement, the main available options are camera-based motion capture or force sensors embedded in the climbing holds. The former could provide a direct measure of body pose, while the latter provide more direct information regarding the load distribution and may potentially convey information that is more difficult to gather by the climbing instructor through direct observation of the climbing exercise. For this reason we opted for force sensors, though the integration of camera-based motion capture was planned as a follow-up and is currently being pursued. A potentially interesting solution was to integrate force and grip pressure sensors in the holds. This, however, would have meant modifying the holds structure: we would no longer have been able to use standard climbing holds, and this would have greatly restricted the variety of scenarios that can be realized at a reasonable cost. Climbing instructors and physiotherapists in particular highlighted the importance of being able to use a wide variety of hold shapes and textures and to easily replace or rotate holds on a climbing route. We, therefore, chose to use force sensors integrated into the wall surface and designed to allow the mounting of standard climbing holds with standard M10 attachment screws. While (21) explored the opportunity of using 1 or 2-Degrees-Of-Freedom (DOF) force sensors to instrument a climbing wall, with the measurement axes lying in the plane parallel to the wall, a full 3-DOF sensor may provide more complete information regarding the pose of the climber’s body. We, therefore, decided to follow the route of a 3-DOF sensor, with a complete redesign of the sensor geometry to limit costs and manufacturing complexity.

Our need-finding session also highlighted how a wall might be used by novice and expert climbers, with an estimated body weight ranging from 20kg for younger kids to 80 kg, which is the reference weight of an adult climber in the EN norms on climbing equipment. Physiotherapist also highlighted the potential interest in using the wall during therapy for children with mild motor disabilities, in particular hemiplegia. For this, sensors would need to resolve the force generated by the child’s affected side. We estimated that a resolution of 5 N would allow resolving the force generated by the arm of a child passively posed on a hold. The measurement range thus should go from 5 N, to over 1600 N, assuming a factor 2 in a dynamic load of an 80 kg climber. Note that, like the 80 kg climber weight, the factor 2 was chosen in accordance with the parameters used to define the structural integrity requirements of climbing holds in EN 12572-3. Because of the need to use the sensors in several exercises and to simplify reconfiguration, the sensor design should also not need recalibration when holds are changed, providing measurements that are independent of the hold shape and size. Due to the variety of commercially available hold shapes, this led us to formulate a requirement that the sensing device should provide a measure of the force vector that is as much as possible insensitive to the point of application of the force on the hold. Finally, in order to be able to compose data from multiple sensors, the force signals should be synchronized and referred to a common reference frame.

Concerning software requirements, an important aspect that emerged from the needs expressed by all the experts was to provide real-time access to the collected data, together with the possibility to visualize it and perform subsequent statistical analysis. The measured signals should be, therefore, visualizable by the climber, climbing instructor, or physiotherapist, on a portable device (smartphone, tablet, laptop), through an interface that is simple enough to be used during or immediately after the exercise without technical training. Information gathered through the sensors should also be represented so as to be understandable by different users. As an example, the climbing instructors might be interested in storing and analyzing the numerical load information later. The climber might instead be interested in a graphical, simplified, and real-time representation of the same information. In both cases, referring signals to a graphical representation of the wall was considered to aid interpretation. We also found a preference for expressing forces in kgf as opposed to N, which allowed simpler considerations regarding body weight distribution. Finally, in order to comply with the need to reconfigure the wall easily, data visualization and sensors-related aspects of the application interface should easily adapt to a change in sensor position, the addition or removal of some sensors, or the repositioning of the holds on the wall. Table 1 summarizes the main requirements that emerged through this process.

At the time of implementing the design, we had the chance of installing the sensorized wall in a public sporting center that caters mostly to children aged 6 to 13yo. As a consequence, we decided to size the structure for these users. The structure is composed of independent modules (3 side-by-side, in Figure 1), each 1200 mm wide and 2500 mm high, supported by a steel frame that allows a variable inclination of the wall of about 15∘ either side from the vertical. The wall could of course be realized with taller modules up to the maximum height of 4.5 m allowed by norm EN 12572-2 (37) with minimal changes to the design.

The steel frame, the 20 mm-thick plywood covering, and all structural details were designed according to EN 12572-2. This allowed us to open the wall to public use. The structure was designed for outdoor use, and a cover was added to protect the sensors from rain and to eliminate the possibility of young users climbing over the wall.

In the wall in Figure 1, the central module is sensorized, while the two side modules are not. Independently of whether a sensorized or non-sensorized panel is mounted, the distance between hold attachments is 230 mm horizontally and 180 mm vertically, and holds are connected to the wall (or the sensors) through standard M10 screws: the sensor was specifically designed to provide standard M10 attachment for commercial holds and to provide a sufficient frontal surface to fix holds up to 200 mm in diameter securely.

To minimize noise and allow easy reconfiguration of the sensors, we equipped each sensor with a dedicated data acquisition board, itself custom-designed and equipped with a 6-channel analog to digital converter with 24-bit resolution and a sampling frequency of 80 Hz (based on the AD7124-4 chip), and a CAN bus interface (MCP2515). All sensors, therefore, behave like independent nodes of a CAN network and can be mounted and removed without any need to modify the hardware or software configuration of the other sensors.

To minimize the production cost of the 3-DOF sensor we redesigned the sensing element from scratch. The shape, obtained from 20 mm thick steel slabs, was designed to require a single cut on a 2-DOF machine while providing easy access to the surfaces that host the strain gauges, to simplify the process of applying the strain gauges and electronics. The resulting design, which is patent-pending, is reported in Figure 2. The shape of the sensor and the placement of the strain gauges were optimized through FEM analysis to provide good resolution over the range of ∼5 to 1600 N and to withstand loads of up to 2400 N without damage. This range was obtained by assuming the maximum dynamic load of 1600 N, multiplied by a 1.5 safety factor. While the sensor bends slightly under load, FEM analysis predicted that during normal use (i.e., under loads of up to 1600 N) the hold would move by less than 2 mm. The housing of the sensors in the frontal panel was therefore designed to allow for this displacement without contact between the sensing element and the wall. The wooden disk, 80 mm in diameter, protrudes 2 mm from the wall to avoid contact between the climbing hold and the wall under the expected load range provides the surface supporting the hold, and hosts a standard M10 fixing in the center. These disks are clearly visible in the central panel in the wall of Figure 1. Notice that, even though in principle high-level climbers could use the 2mm step as an additional grip, during our test this did not appear to be the case. Note that to mount or move a sensor the user must access the back of the wall. The mounting procedure is quite simple: the sensor is attached through 4 screws, each sensor is equipped with its own acquisition board, and power and data are brought to the acquisition board through a 4-wire, hot-swappable, daisy chain connection. Moving or adding a sensor thus requires attaching 4 screws and connecting the sensor as an arbitrary segment of the daisy chain. When it came to building the sensors, the total cost for the low-volume industrial production of a batch of sensors by a load-cell manufacturer, plus the cost to have the custom-made acquisition boards assembled by a PCB manufacturer was about 310C€ per sensor, in line with the low-cost (and 2-DOF) implementation suggested in (21). These features satisfy (R1)–(R3).

The goal of the software interface is to provide a structured tool to quantitatively evaluate the performance of the climber in real-time and to easily record and export data for subsequent analysis with other software and to simplify the reconfiguration of the sensorized wall, according to (R4)–(R6).

From the point of view of the software interface, the climbing wall was set up as a server that provides data coming from the sensors, while the software interface itself is a mobile application running on a mobile device. The server implements the data collection from the sensors connected through the CAN bus. Data is then made available through a Representational State Transfer (REST) web service and a set of threads and queues. At any given time, a client can connect to the REST service and download the latest data coming from the sensors. The client side of the application is represented by a mobile application running on a modern tablet. In our setup, we used an Apple iPad in two different versions (Apple iPad Pro 10.2″ 2016, and Apple iPad Pro 13″ 2020), running iOS 15. The client app was however written in React Native to allow a simple porting to Android, and has been successfully tested on Android 12. We defined this component as a REST service on the mobile side since it was not possible to define a streaming API to retrieve data. For this reason, the mobile application regularly polls the REST web service to retrieve the data when the application is in recording mode, and the REST web service packs the available data in a JSON structure before returning it to the caller.

On the client side, the mobile application provides four main environments: •

a Setup environment used to configure the sensor network when sensors or holds placement is changed,

The Setup environment, shown in Figure 3, allows to define the parameters needed to connect a tablet to the acquisition system (panel A), to define the number of possible hold placements on the sensorized wall, which are arranged on a regular grid (panel B), to associate sensors and holds with specific placements (panel C), and to define the group of holds that constitute a given climbing problem, attaching optional tags to specific holds (panel D). In particular, when associating sensors to placements, the Setup environment allows to rotate of the sensor reference frame (Figure 3C) so that the reference frames of all sensors are aligned, with the x axis parallel to the horizontal edge of the wall and pointing right, the y axis parallel to the vertical edge and pointing up, the z axis pointing out of the wall. This allows having a coherent framework for all sensors, which is necessary when composing forces measured on different holds. After having assigned a hold to a placement on the grid, the user can select a visual representation of the hold from a menu of pre-defined shapes (this menu and the corresponding configuration step are not displayed in the figure). This helps to more easily relate the wall virtual representation, visible in Figures 3 and 5, with the real wall, visible in Figure 7. Furthermore, as mentioned before, the Setup environment allows attaching one or more user-defined tags to the sensors (Figure 3D), to automate subsequent classification and analysis in the Data Management and Recording Analysis environment. In the tests that were performed for this study, since we had dedicated distinct sets of holds to hands and feet, we tagged as Hands the upper 5 holds and Feet the bottom five ones (see Figure 3, bottom right panel).

The User Detail environment, in Figure 4, is used to insert user data that will be attached to a recording. In particular, we decided to include first and last names, age, height, and weight, the last two being relevant to parametrize many biomechanical models. The weight is also used in the Data Management and Recording Analysis environment (Figure 6B), to visually compare the total force expressed on tagged holds with the climber’s body weight.

The Data Acquisition environment, in Figure 5, is used to handle data recording. It shows a schematic representation of the climbing wall with the holds, positioned as defined through the Setup environment, with an overlay reporting the measured forces in real-time, and it allows to start and stop a recording and assign it a name.

The Data Management and Recording Analysis environment, in Figure 6, proposes different views of the recorded data and allows to export it in JSON format. Panel A shows the Data Management menu, where recordings can be selected for analysis, exported, or deleted. The other three panels show three preview modes. Implemented preview modes allow to select an arbitrary lowpass filter bandwidth, to visualize the magnitude of the vector sum of all forces measured on holds with a given tag (in our example: Feet and Hands, panel B), to visualize the magnitude of the force measured on each hold (panel C), and to visualize the three components of the force measured on each hold (panel D, x in blue, y in red, z in green). In the last two modes, the user can tap on a hold in the visual representation of the grid, to highlight the corresponding set of plots. In the screenshot, the central hold, in position x=3, y=8, has been tapped and is highlighted in red, together with the corresponding plots. Note that, in all these screenshots, forces are expressed in kgf according to the experts’ preference. Data is however recorded in N, and the analyses that we performed on the exported data and report in later sections utilize SI units.

To bring client and server components together, we decided to leverage a local WiFi network where both client and server are connected. We opted for WiFi instead of Bluetooth as this increases the possibilities in terms of deployment of the application and does not require executing native code that needs to be signed and distributed through Apple’s official channels (both for production, development, and testing of the application). Such a local network can, but does not have to, be connected to the internet.

In order to test the functionality of the hardware and software setup in a realistic data acquisition session, we recruited a total of 11 adult climbers. The study was approved by the Ethics Committee of Politecnico di Milano, and all participants read and completed an informed consent form. We report in Table 2 their declared years of climbing experience, and maximum redpoint grade reached in the 3 years prior to the experiment, in the French/sport scale (see e.g., (38) for a description climbing grade scales). Climbers 5 and 6, classified as novice in the table, were at their first experience on a climbing wall.

We initially characterized the force sensor performances on a bench-top hydraulic tensile/compressive testing rig (MTS 858 Mini Bionix II) to evaluate sensor linearity, hysteresis, and crosstalk, for loads directed along the sensor’s x, y, z axis. A single sensor was tested. Force sensors are required to measure the force vector irrespective of the point of application on the hold, that is, irrespective of eccentricity. To test this, we loaded the sensor in the x and y direction with an offset of 15 and 90 mm along the z axis from the hold attachment point, and in the z direction with an offset of 0 and 80 mm along the y axis (for definition of coordinate system, see Figure 2). The results are representative of the expected performance when using holds of diameter up to roughly 180 and 58 mm thick, assuming the wall thickness to be 20 mm, plus the 2 mm spacer. All tests were performed with nominal loads FN ranging from 0 to 1600 N in 13 steps (0, 320, 640, 960, 1280, 1600, 1280, 960, 640, 320, 0 N), covering the range of loads identified in the requirements.

Then, a total of 10 sensors (mechanical component and acquisition board) were assembled and networked (only 10 sensors were available at the time of this testing). The network of sensors was mounted on a wall (see Figure 7) with a front panel adapted to host the 10 sensors in any one of 65 possible locations, arranged in a 5 by 13 grid. The sensorless locations were covered with a wooden lid equipped with standard M10 inserts, though for the purpose of our testing, only the sensorized locations were utilized. Once the sensors were on the wall, the APP setup time for the 10 sensors was less than 10 min.

The sensorized wall was then used as a platform for the climbing exercise, during which data was acquired. All climbers were instructed to perform a short warm-up session on the ground before starting the climbing exercise. The exercise consisted of a circular circuit on the sensorized wall with a fixed sequence of handholds, as depicted in Figure 7: climbers started with right hand on 1R and left hand on 1L and had to move right and left hand alternatively through the sequence 2R-2L-3R-3L-1R-1L, which brought them back to the starting position. Feet placements were not fixed to allow climbers to optimize their moves as they practiced the circuit. The wall angle was set at 90∘. Climbers had no prior experience with the circuit before the test and were instructed to repeat the circular circuit as long as they could endure.

The results of the characterization of the sensor are reported in Tables 3 and 4.

One further performance metric regards the noise characteristic of the sensor. As mentioned before, the force sensors sample at 80 Hz. Without further filtering, during lab acquisitions, we measured a Gaussian noise (Kolmogorov Smirnov one-sample test) with a standard deviation less than 1 N. This suggests that the sensors will be able to resolve load signals of amplitude well below 5 N with bandwidth below 40 Hz, which is where we expect to find most of the information generated by human motion.

All climbers reported that the slight movement of the holds due to the sensor deformation under load, and the gap between the hold and the wall, were not perceivable during the climb. The acquisition app collected the simultaneous data streams from the 10 sensors, sampled at 80 Hz, without data loss, as expected. The drift in the time stamps attached to each data point by the acquisition boards with respect to an external clock was not detectable, throughout the length of the experiment. Measures returned by each hold allowed to clearly discriminate the direction and intensity of the force in relation to the task and the position of the hold. See, as an example, Figure 8, relative to the trial with Climber 4, which lasted about 9 min. Here, hold 2L was mostly loaded along the vertical axis, while holds 2R and 3R, which were used to move up and down between 1R/3L and 2L, respectively, show a significant Fx-component, as one would expect to balance the weight transition between feet. Reference system alignment and prior tagging allowed for simple post-processing of the data. In our test (see Figure 9) this allowed us to compute the magnitudes |Fhands(t)| and |Ffeet(t)| of the vector sum of the forces on all holds tagged Hands and Feet at time t, to discriminate the contribution of upper and lower limbs. A similar process was followed for Figure 10, which displays the hands to total force ratio R(t) , for the 11 test climbers, defined as(1)R(t)=Fhands,avg(t)/(Fhands,avg(t)+Ffeet,avg(t)).In the above formula, Fhands,avg(t) and Fhands,avg(t) are the moving average, over a 60s window, of |Fhands| and |Ffeet|, respectively.