2.1. Olfactory robotics

Although often considered pollutants, odors in compounds provide significant benefits for environmental monitoring, particularly in gas leak detection, where the human olfactory system alerts individuals before they are exposed to hazardous levels of harmful substances [Reference Zanchettin, Almeida and de Menezes44]. However, this detection procedure must not be performed by humans through smell, as it can lead to various health issues, such as dizziness or long-term damage to the olfactory nerves due to toxic exposure [Reference Jalal, Alam, Roychoudhury, Umasankar, Pala and Bhansali45]. Additionally, in confined environments such as gas lines and scenarios with different volatile organic compounds (VOCs), the human smell is not sufficient. This makes assistive technologies essential to identify and access odor sources. To address this challenge, researchers have turned to biomimicry and bio-inspired engineering – i.e., the development of olfactory tools mimicking the sensing abilities of animals such as dogs and insects [Reference Pomerantz, Blachman-Braun, Galnares-Olalde, Berebichez-Fridman and Capurso-García46, Reference Li, Covington, Tian, Wu, Liu and Hu47]. These tools, known as e-noses, function as artificial odor sensors and are typically mounted on stationary structures or mobile robots [Reference Cheng, Meng, Lilienthal and Qi48, Reference Heredia, Cruz, Balbin and Chung49] to collect static and dynamic data, respectively.

Researchers have explored various applications of static and mobile olfactory robotic systems. For instance, Gonzalez et al. [Reference Gonzalez-Jimenez, Monroy and Blanco50] developed the multi-chamber electronic nose to address the long recovery times typically required by metal oxide semiconductor technology after gas exposure. Maho et al. [Reference Maho, Dolcinotti, Livache, Herrier, Andreev, Comon and Barthelme51] designed an optoelectronic olfactory robotic system for long-term gas discrimination. Harianto et al. [Reference Harianto, Rivai and Purwanto52] implemented an e-nose on an omnidirectional robot, while Abdelkhalek et al. [Reference Abdelkhalek, Alfayad, Benouezdou, Fayek and Chassagne6] introduced a compact, embedded e-nose for both volatile and non-volatile odor classification in robotics. Xing et al. [Reference Xing, Vincent, Fan, Schaffernicht, Bennetts, Lilienthal, Cole and Gardner53] and Fan et al. [Reference Fan, Bennetts, Schaffernicht and Lilienthal54] demonstrated gas discrimination and mapping capabilities for emergency response scenarios using mobile robots, including the FireNose system, which can navigate harsh environments. A study by Pique et al. [Reference Piqué, Stella, Hughes, Falotico and Santina55] explored a smell-driven soft robotic arm for navigation, and Burgues et al. [Reference Burgués, Valdez and Marco56] presented a high-bandwidth e-nose for rapid tracking of turbulent plumes. Liu et al. [Reference Liu, Zhang, Wang, Qin, Yang, Sun, Hao, Shu, Liu, Chen, Zhang and Tao57] developed a star-nose-inspired tactile-olfactory bionic sensing array for robust object recognition in non-visual environments, and Sanchez et al. [Reference Sanchez-Garrido, Monroy and Gonzalez-Jimenez58] introduced a configurable smart e-nose for spatio-temporal olfactory analysis. Collectively, these studies offer design concepts and prototypes that demonstrate how robotic mobility and agility can effectively enhance the transport and deployment of e-noses for both domestic and field applications.

2.2. Odor classification in olfactory robots

Odor classification is the technique of recognizing and classifying various odors according to their chemical makeup and sensory quality [Reference Rasekh, Karami, Wilson and Gancarz59]. In nature, odors are picked up by specialized receptors reacting to VOCs, enabling humans to discern between different scents [Reference Sankaran, Khot and Panigrahi60]. These specialized receptors are mimicked by olfactory sensors and e-noses, which are often deployed by robotic systems, such as the one shown in Figure 2.

Figure 2. Conceptual framework. (a) CAD rendering (Left) of the olfactory SGR system demonstrating one-dimension (top) and two-dimensional growth (bottom) toward a target odor source. A full eversion cycle occurs when the robot grows and retracts back to its initial position in the base. The e-noses are transported through a tip-mounted mechanism (see Figure 4 for details). (Right) Olfactory process: The robot uses four e-noses (TGS 2600, 2602, 2611, and 2220) to collect raw data in both static and transit readings. The collected data are then processed into a machine-learning pipeline for autonomous gas classification tasks in SGR systems. We allocated $60 \%$ of this data for training twelve different ML models, $20 \%$ for testing, and $20 \%$ for validation. The test results show that the k-Nearest Neighbors (KNN) classifier outperformed all other models with an accuracy of $99.88 \%$ .

To boost odor classification of olfactory robots, either in static or mobile applications, researchers have used several algorithms and ML models. For instance, support vector machine (SVM) has been used to aid the classification of three different VOCs; ethanol, methanol, and acetone, using three TGS sensors (2,600, 2,602, and 2,620) from a static position, with the robot achieving $95 \%$ accuracy [Reference Nyayu, Handayani, Nurmaini and Yani61]. Another study used a metal oxide semiconductor e-nose and a KNN classification algorithm to differentiate between four fruits and juices on a static robot [Reference Abdelkhalek, Alfayad, Benouezdou, Fayek and Chassagne6]. This study achieved a classification accuracy of $98 \%$ . Yan et al. [Reference Yan, Cao, Zhang and Yang62] evaluated a fuzzy linguistic odor cognition method for robotic olfaction. This approach centers on realizing a knowledge representation inspired by the brain using fuzzy language rules, which outperforms methods like SVM in uncontrolled, realistic conditions. Furthermore, continuous time recurrent neural networks built on evolutionary robotics were presented by De et al. [Reference de Croon, O’connor, Nicol and Izzo15] using a simulated robot with chemical and wind direction e-noses. The approach proves effective for odor-source localization. Li et al. [Reference Li, Feng, Ge, Zhu and Zhao63], on the other hand, utilized an ensemble learning method found to enhance active perception in robot electronic noses. Additional approaches to odor classification in olfactory robots include deep learning with feature extraction, a hybrid neuromorphic-bayesian model, and localization using vision and fusion navigation algorithms [Reference Hassan, Wang and Mahmud64–Reference Kausar, Zayer, Viegas and Dias66].

However, in contrast to the static odor detection problem, odor classification in a mobile olfactory system (e-nose installed on a moving platform) introduces open challenges, mostly because existing e-noses show a notable delay in both response and recovery times to gas exposure, rather than responding to it instantly. Due to this delay, and depending on the type of sensors, e-noses are often unable to achieve a steady state or provide a signal that is powerful enough to correctly identify the odor/gas being detected. This is still an open challenge [Reference Trincavelli, Coradeschi and Loutfi8, Reference Monroy and Gonzalez-Jimenez67], while most olfactory robotic studies on odor classification are mostly based on pattern recognition in static or grounded robots.

2.3. Softgrowing robot applications

Most SGRs are a class of bio-inspired mechanical systems similar to vine plants in their growth-trailing behavior. The primary growing mechanism involves internal air pressure pushing material through the center of the robot’s flexible, tubular body, turning it inside out at the tip in a process known as eversion. One advantage of this actuation process is that the robot’s body extends forward without dragging or sliding against the environment, allowing it to access confined spaces with minimal friction [Reference Hawkes, Blumenschein, Greer and Okamura68].

Recently, the field of SGR has gained research interest in modeling, computational design, control, and applications. Models such as quasi-static kinematics have been applied to account for the robot’s length, curvature, and bending stiffness to accurately predict growth and navigation parameters [Reference Tutcu, Baydere, Talas and Samur69]. Motion planning, teleoperation, and vision-based autonomous steering have also been employed for smooth navigation during SGR operation [Reference Coad, Blumenschein, Cutler, Zepeda, Naclerio, El-Hussieny, Mehmood, Ryu, Hawkes and Okamura25, Reference Stroppa, Luo, Yoshida, Coad, Blumenschein and Okamura37, Reference Wu, Hadi Sadati, Rhode and Bergeles70]. This class of robots has proven its potential through various indoor and field demonstrations. Examples include navigating unknown environments using passive deformation [Reference Fuentes and Blumenschein71], deploying distributed sensor networks to relay directional temperature and humidity data in hard-to-access spaces [Reference Gruebele, Zerbe, Coad, Okamura and Cutkosky72], exploring underground tunnels at archeological sites [Reference Coad, Blumenschein, Cutler, Zepeda, Naclerio, El-Hussieny, Mehmood, Ryu, Hawkes and Okamura25], investigating coral reefs [Reference Luong, Glick, Ong, DeVries, Sandin, Hawkes and Tolley73], delivering tools and life support to trapped victims [Reference Jeong, Coad, Blumenschein, Luo, Mehmood, Kim, Okamura and Ryu32], operating as part of a cyber-physical measurement system [Reference Grazioso, Tedesco, Selvaggio, Debei and Chiodini74], and clinical applications [Reference Oyejide, Stroppa and Saraç75].

Although research on olfactory robotics continues to grow, to the best of our knowledge, no prior studies have specifically investigated odor identification and classification using SGRs. This is the first instance of a SGR being used for odor detection and the use of the collected data in a ML classification.

In this work, we introduce the concept of an olfactory SGR. SGRs possess unique advantages for olfactory tasks: they can elongate to significant lengths with minimal frictional interaction with their surroundings, they operate with fewer actuators than conventional mobile platforms, and they enable inherently safe, non-destructive exploration in confined or sensitive environments. They are also lightweight, moderately sized, and fabricated from low-cost, readily available materials [Reference Coad, Blumenschein, Cutler, Zepeda, Naclerio, El-Hussieny, Mehmood, Ryu, Hawkes and Okamura25, Reference Blumenschein, Coad, Haggerty, Okamura and Hawkes26], making them more deployable than most rigid static or wheeled robots. The novelty of this work, therefore, lies in leveraging the intrinsic design properties of SGRs; their compliant growing mechanism, low mechanical complexity, and safe interaction characteristics, to enable a new modality of odor detection and classification that conventional robotic platforms cannot achieve.

Table I and Figure 3 provide a comparative description of our proposed olfactory SGR alongside previous olfactory robotic systems, based on key performance and design metrics. The static and wheeled/mobile systems shown in Cases 1 and 2 rely on conventional mechanical actuation and require either reverse traversal or manual extraction for sensor retrieval. This is reflected in their higher actuator demands, greater environmental interaction, and limited adaptability [Reference He and Gao76, Reference He, Wang and Liu77]. Static platforms exhibit restricted workspace coverage, while wheeled robots offer only moderate reach in confined environments due to locomotion constraints and the need to physically navigate the environment.

Table I. Comparison of olfactory robots based on key performance and design metrics.

Figure 3. Case 1 (left): Reverse actuation in static [Reference Moshayedi, Khan, Shuxin, Kuan, Jiandong, Soleimani and Razi5, Reference Gonzalez-Jimenez, Monroy and Blanco50], and mobile [Reference Xing, Vincent, Fan, Schaffernicht, Bennetts, Lilienthal, Cole and Gardner53, Reference Arain, Bennetts, Schaffernicht and Lilienthal78]. For a static olfactory robot equipped with an actuated arm, the arm is simply driven in reverse to retrieve the e-nose after sampling. In mobile platforms, such as wheeled robots, retrieval may require retracing the traversal path to return the sensor to its origin. Case 2 (top and bottom): Physical removal of the system. In some static configurations, the entire robot or sampling assembly must be manually removed to recover the sensor. Likewise, mobile robots may need to be physically carried out of the environment if reverse traversal is impractical or unsafe. Case 3: Softgrowing robots. Inspired by vine-like tip growth, SGRs advance through tip eversion, enabling continuous odor data collection along the growth trajectory [Reference Hawkes, Blumenschein, Greer and Okamura68]. A single SGR can operate in both static and mobile scenarios using electropneumatic actuation for growth and retraction [Reference Coad, Blumenschein, Cutler, Zepeda, Naclerio, El-Hussieny, Mehmood, Ryu, Hawkes and Okamura25], while navigating either passively with minimal environmental contact or actively without requiring direct interaction with surrounding surfaces. This mechanism allows sensor deployment and gas removal [Reference Lee, Kim, Seo, Park and Ryu79] without traditional reverse motion or physical extraction.

In contrast, the SGR in Case 3 shows more workspace coverage with minimal environmental contact, consistent with its tip-eversion mechanism that eliminates the need for backtracking or physical removal. Its low actuator requirement and high energy efficiency arise from pressurized eversion rather than jointed or wheel-based locomotion. Furthermore, the high adaptive morphology possible with SGRs enables reconfiguration and navigation in confined or irregular spaces – capabilities not achievable with rigid grounded or mobile platforms.

In the following sections, we will implement this concept by developing the robotic system and evaluating its performance through static and transit odor detection experiments.

3.1. Robot design for non-contact eversion

This base station is a key component of the SGR system, often composed of a pressurized chamber housing the main spool, which directs the robot’s orientation during eversion. There are two primary configurations for mounting the main spool: (i) vertical mounting of the spool directly onto the motor shaft, as in the design by Coad et al. [Reference Coad, Blumenschein, Cutler, Zepeda, Naclerio, El-Hussieny, Mehmood, Ryu, Hawkes and Okamura25], or (ii) horizontal mounting with actuation via a pulley–belt mechanism [Reference Hofmann, Fankhauser, Willen, Inniger, Klahn, Löffel and Meboldt80]. We opted for the first approach (Figure 1) for two reasons: (i) it enables direct and rapid rotation of the spool with minimal torque from the motor, and (ii) it allows the robot to grow mid-air through a vertical base without contacting its surroundings. This non-contact advantage is particularly relevant when transporting sensors or tools over sensitive environments. Our empirical tests show that the robot can extend up to $150$ cm without collapsing when the tip mount is absent. However, in this study, we limited the operational length to $100$ cm (Figure 5) due to the payload from the tip mount device. The vertical spool orientation also allowed us to characterize the buckling behavior induced by the tip mechanism under the robot’s own weight.

As shown in Figure 1, the base is 3D-printed using polylactic acid and assembled in four parts: (i) the top seal, which houses the pneumatic check valve, connecting pins, and pressure sensor; (ii) the chamber, which stores compressed air and houses the inner spool with its electronic components; (iii) the actuation platform, located at the front of the cylinder, bearing housings for four brushless DC motors, connection circuits, and a small cylinder from which the everted LDPE tube grows; and (iv) the bottom seal. The seals are secured to the cylinder brim using silicone O-rings inserted into grooves, while the flanges are fastened with screws and nuts through $0.8$ cm holes in the cylinder flange to provide reinforcement.

3.2. Fabrication of the soft-growing robot body

We fabricated the inflatable robot body as $8.5$ cm diameter tube using flat LDPE sheets. We selected LDPE because of its low elastic modulus, high elongation at break, weldability, and suitability for pressure-driven eversion [Reference Blumenschein, Coad, Haggerty, Okamura and Hawkes26]. As described in Figure 5, we cut the flat sheet to a width of $31$ cm and marked $2$ cm sealing margins along both longitudinal edges, leaving a $27$ cm central active region for inflation. To enable controllable tip bending, we integrated two symmetric pulling cables onto the sheet. We positioned the cables at a distance of $6.57$ cm from each $2$ cm sealing margin. With this configuration, the robot bends right or left when a differential compressive load is applied at the tip or when distributed loading accumulates along one side of the body. The cables also support the tip mount during eversion.

We then routed the cables through polytetrafluoroethylene stoppers measuring $1$ cm in length and $0.4$  cm in diameter, spaced at $5$  cm intervals. We secured each stopper to the LDPE sheet using double-sided adhesive tape to maintain precise cable alignment during eversion.

To form the pressurizable tube, we folded the sheet lengthwise onto itself with all stoppers enclosed and sealed the full length along the $2$ cm margins using an impulse heat sealer. We also sealed the distal end of the tube, leaving a $0.6$ cm central barrier to accommodate the folded sensor wires and the soft aspirator tube. After sealing, we removed the barrier, inserted the wires and tubing, and wrapped the tail tightly with tape (Gorilla weatherproof adhesive tape) to prevent air backflow during pressurization.

3.3. Tip mount design, mechanism, and evaluation

The material at the tip of SGR continuously changes as it grows and retracts; therefore, maintaining a tip mount that remains steadily fixed on the end effector is a significant challenge. Several designs in the literature have addressed this issue [Reference Stroppa, Selvaggio, Agharese, Luo, Blumenschein, Hawkes and Okamura38, Reference Luong, Glick, Ong, DeVries, Sandin, Hawkes and Tolley73, Reference Greer, Morimoto, Okamura and Hawkes81]. However, previous designs require embedding the entire tip mount drive system (whether motors or magnets) within the device, which limits optimal performance in non-contact eversion, particularly during mid-air operation. Motivated by this challenge, we developed a simplified tip-driven mechanism that relocates the drive components to the robot’s base. This relocation not only reduces the payload at the tip but also improves growth dynamics and simplifies assembly, while preserving the stability of the tip mount during eversion and retraction. Figure 4 provides a detailed description of our tip mount mechanism, while Sections 3.3.1–3.3.4 present a characterization of its effects on the robot’s eversion and navigation.

3.3.1. Growth and retraction modeling of the tip-mounted eversible robot

To describe the kinematics of our SGR with the tip-mounted assembly, we first model the interaction between pneumatic eversion and the mechanical resistance introduced by the roller-based mount and side-tension cable system. The geometric sketch is illustrated in Figure 6(a). During growth, the LDPE robot body everts outward through the tip, driven by internal pressure $P$ . The net force applied to the interior surface of the tip mount is calculated from Eq. (1):

(1) \begin{equation} F_{\text{tip}} = P \cdot A, \ \text{;} \ A = \pi r_i^2 \end{equation}

where $A$ is the robot’s area and $r$ the radius.

For successful eversion, this force must exceed the combined resisting force $F_f$ due to friction at the roller interface as given in Eq. (2):

(2) \begin{equation} F_{\text{tip}} \gt F_f = \mu _r \cdot N \end{equation}

where $\mu _r$ is the coefficient of friction between the Dycem-covered rollers and the robot body, and $N$ is the normal preload applied by the extension springs. Once the frictional threshold is exceeded, the robot extends to length $L$ at a rate governed by the supplied airflow over time $ \dot {V}(t)$ , with the resulting tip displacement given by Eq. (3):

(3) \begin{equation} \frac {dL(t)}{dt} = \frac {\dot {V}(t)}{\pi r_i^2} \end{equation}

As the robot continues to grow, the spring-loaded rollers conform passively to the robot’s surface, maintaining grip while allowing smooth extension until the system approaches its burst pressure. The tip remains mechanically constrained within the rigid cap, generating a localized tension across the mount length. During retraction, the internal pressure is lowered by a pressure relief valve, and the motorized spool $M_1$ actively back-drives the body. Retraction occurs when the applied motor torque $\tau$ from $M_1$ overcomes the combined resistive torques from the robot’s body tension and roller friction in Eq. (4):

(4) \begin{equation} \tau \geq r_s \cdot (T + F_f) \end{equation}

where $r_s$ is the spool radius, and $T$ is the longitudinal tension generated in the body and cables during retraction. The symmetric routing of the side cables ensures uniform retrieval unless differential actuation is introduced by motors $M_2$ or $M_3$ .

To characterize the bending behavior of the robot under different mechanical configurations, we compare two distinct cases presented in Sections 3.3.2 and 3.3.3: (i) a segmented robot with selectively stiffened links (e.g., via inflated circumferential pouches) that localize deformation at discrete joints, and (ii) a continuously soft, pressurized beam without local stiffness modulation.

Figure 6. Cross sections of the tip-mounted eversible robot. (a) Side axial cross-section of the robot. The robot body comprises a pressurized LDPE body, which resists deformation due to pressure-induced stiffness. Two cables (braided spectra lines) of lengths $L_R$ on the right, and $L_L$ on the left, run symmetrically along the robot body from external spools $M_2$ and $M_3$ to grooves at the rear of the tip mount. (b) Circumferential cross-section. The cables are routed through stoppers, which keep them aligned and create a contraction limit during bending. (c) Bending through active link stiffness (detailed in Section 3.3.2). The bending motion of the tip-mounted eversible robot is achieved through differential actuation, exploiting the geometry of the tip mount and the cables. When one cable is pulled while the other is released or held, asymmetric tension induces a curvature toward the tensed side. (d) Constant-curvature model of continuum-like robots.

3.3.2. Stiffened link with localized bending

When SGRs are steered from the tip, the geometry is typically a curved segment with buckling at the base [Reference El-Hussieny, Hameed and Ryu82] (Figure 6(d)). One way to achieve selectively stiffened links and active segments is through layer jamming in a circumferential pouch [Reference Exarchos, Wang, Do, Stroppa, Coad, Okamura and Liu30, Reference Do, Banashek and Okamura34]. To model the mechanical behavior of our tip-mounted eversible robot for active bending without buckling at the base and desired link, we considered a nonlinear formulation that captures the gradual resistance introduced by the stoppers when the robot’s tip is pulled.

As the robot bends, the entire length changes by $\Delta l_o$ , while the pulled cable shortens by a differential amount $\Delta l$ (Figure 6(c) and (d)), and the contraction ratio is defined in Eq. (5) as:

(5) \begin{equation} a = \frac {\Delta l}{l_0} \end{equation}

where $l_0$ is the original length of the bending section. In the early stages of bending, adjacent stoppers do not make direct contact, and the system permits relatively free deformation. However, as bending increases, the stoppers begin to contact one another at a critical threshold defined by Eq. (6) [Reference Jitosho, Simón-Trench, Okamura and Do83]:

(6) \begin{equation} a_0 = \frac {l_g}{l_s + l_g} \end{equation}

where $l_s$ and $l_g$ represent the lengths of the stopper body and inter-stopper gap, respectively. Beyond this threshold, the stoppers compress against one another, introducing a nonlinear restoring force modeled in Eq. (7) as:

(7) \begin{equation} F_s(\Delta l) = \begin{cases} 0, & \Delta l \lt a_0 l_0 \\ k_s (\Delta l - a_0 l_0)^n, & \Delta l \geq a_0 L_0 \end{cases} \end{equation}

where $k_s$ is the stiffness coefficient associated with the stopper material, and $n$ is the empirical nonlinearity exponent. This stopper force contributes to the total torque that must be overcome for bending. To achieve or maintain the bend angle $\theta _b$ , we estimate the tension needed to hold or reach a specific bend angle $\theta _b$ , given a known internal pressure $P$ and robot tube geometry (i.e., moment arm $r$ ).

Assuming the pulling cable exerts a tension $T$ with an effective moment arm $r$ , and the internal pressure $P$ induces a resisting torque $\tau _P$ , the total torque balance (considering Eqs. (7) and (8)), is:

(8) \begin{equation} T = \frac {\tau _P}{r} + F_s(\Delta l) \end{equation}

Intuitively, from Eqs. (1) and (2), bending is initially governed by internal pressure and cable tension alone. However, in our design, the mechanical constraint imposed by the stoppers rapidly increases as the critical bend is approached, thereby defining the robot’s maximum curvature. The nonlinear model provides a more realistic estimate of actuator force requirements for our tip-mounted eversible robot, and can be adjusted to match experimental bending force profiles by tuning $k_s$ and $n$ .

3.3.3. Non-segmented continuous bending

Non-segmented SGRs typically assume a constant-curvature model when actuated from the tip. Their behavior, both with tip-mounted steering systems [Reference Jitosho, Simón-Trench, Okamura and Do83, Reference Coad, Thomasson, Blumenschein, Usevitch, Hawkes and Okamura84] and without [Reference Blumenschein, Coad, Haggerty, Okamura and Hawkes26, Reference El-Hussieny, Hameed and Ryu82], has been extensively studied in prior work. A second case in our model shows the mechanical behavior for continuous bending with the spatially spaced stoppers and dual pulling cables along the robot body (Figure 6(d)).

Upon actuation, a tension $T$ is applied to one of the cables, initiating bending from the tip. Unlike in the stiffened-link model (Section 3.3.2), this tension causes distributed curvature along the robot’s length due to its compliant, pressurized structure. The stoppers act as constraint points, limiting relative displacement between adjacent points on the body. As the pulling continues, stoppers sequentially engage, accumulating local strain until adjacent stoppers at the joint reach the critical bending limit.

While the resistance torque still originates from the internal pressure $P$ , it is now distributed along the body. The effective moment arm $\tilde {r}$ , shorter than the robot’s outer radius $r$ , depends on the curvature distribution and local deformation shape. The tension required to achieve a given bend angle $\theta _b$ under this configuration can be approximated by Eq. (9) [Reference Nesler, Swift and Rouse85]:

(9) \begin{equation} T = \frac {|\tau _P|}{\tilde {r}} = \pi \tilde {r}^2 P \left ( \tan ^2 \left ( \frac {\theta _b}{2} \right ) + 1 \right ) \end{equation}

From Eq. (9), and when compared to Eq. (8), it is obvious that a larger tension is required to achieve the same bend angle compared to the segmented robot in Figure 6(c) and Section 3.3.2. Nonetheless, the stopper–cable system acts to constrain and guide the deformation, enabling more predictable bending even in the fully non-segmented configuration.

3.3.4. Structural load analysis of tip mount assembly

The tip mount was designed to maintain a dynamic balance of forces during internal pressurization and actuation. The working pressure was empirically determined to ensure structural integrity while enabling mid-air eversion. It needed to be sufficient to both clamp the robot body to the rollers and support the tip-mounted device without inducing collapse. We found a pressure of $8.8$ kPa to provide a reliable balance between these requirements. Upon inflation of the LDPE robot body, this internal pressure $P$ produces a radially outward force against the inner tube wall, pressing it firmly against the rollers. This clamping effect increases the normal load $N$ at the roller interface, thereby enhancing friction and improving grip between the robot body and the mount.

Given that the rollers have a diameter of $1.6$ cm and a width of $3.8$ cm, and that the tip mount assembly has a total mass of $86$ g (with a length approximately $14\%$ of the robot’s fully extended length), the normal force was estimated using Eq. (1). Assuming $A$ represents the effective contact area between the robot body and the rollers, the resulting radial force was computed as $F_{\text{rad}} = 1.77$ N.

In addition, we evaluated the structural integrity of the inflated LDPE tube under the gravitational load of the tip-mounted assembly by comparing the axial load $W$ with the critical Euler buckling load $P_{\text{cr}}$ for a fixed–free cantilevered column (Eq. 10):

(10) \begin{equation} P_{\text{cr}} = \frac {\pi ^2 E I}{(K L)^2} \end{equation}

The parameters used were Young’s modulus of LDPE $E = 200\,\text{MPa}$ ; area moment of inertia computed from the hollow circular section with outer and inner diameters $D_o = 8.5$ cm and $D_i = 8.0$ cm, respectively; unsupported length $L = 100$ cm; and effective length factor $K = 2$ for a fixed–free column. The moment of inertia was obtained from Eq. (11):

(11) \begin{equation} I = \frac {\pi }{64}\left (D_o^4 - D_i^4\right ) \end{equation}

yielding $I = 5.52\times 10^{-7}\,\text{m}^4$ for the specified geometry. Substituting these values into Eq. (10) results in a critical buckling load of $P_{\text{cr}} \approx 2.72\times 10^{2}\,\text{N}.$

Lastly, the axial gravitational load exerted by the mounted device was computed using Eq. (12):

(12) \begin{equation} W = m \cdot g \end{equation}

where $m = 0.086$ kg is the mass of the assembly and $g = 9.81\,\text{m/s}^2$ . This yields a downward load of $W = 0.84$ N. Since $W \ll P_{\text{cr}}$ (i.e., $0.84$ n $\ll 272$ N), the tube exhibits an extremely high safety margin against global Euler buckling under the applied load. It is worth noting that this Euler estimate is conservative: internal inflation pressure typically increases axial stability (pressure-stiffening), further enhancing resistance to buckling beyond the classical solid-column prediction.

3.4. Airflow through passive aspirator

In our olfactory SGR system, olfaction is enabled via a passive aspirator mounted at the extreme of the robot’s tip mount (Figures 4 and 7). The aspirator inlet has a cross-sectional area $A_{\text{in}} = 1\,\text{cm}^2$ , a length of $1.2\,\text{cm}$ , and operates under ambient room temperature ( $T \approx 25^{\circ }\text{C}$ ). It functions as a passive inlet, driven by pressure gradients generated by ambient airflow induced by a portable $5$ V fan at maximum speed $v_{\text{fan}}$ , and the robot’s translational motion. We chose a single, well-defined inlet geometry to ensure that the passive aspirator entrains VOCs in a consistent flow pattern, thereby promoting uniform sample acquisition and enabling the generation of repeatable data across all exposure trials.

Under steady, incompressible flow conditions, the volumetric flow rate into the sensor chamber is approximated as (Eq. 13):

(13) \begin{equation} Q_{\text{asp}} = C_{d} \, A_{\text{in}} \, v_{\text{eff}} \end{equation}

Figure 7. (a) Side view of the main geometry. The passive aspirator features two inverted inlets and outlets for aspiration and relief. The aspiration inlet aligns with the floor of the connection platform and extends $1.2$ cm into the sensor chamber. The relief inlet connects from inside the chamber to an outlet protruding $1.5$ cm above the connection platform, sealed with a one-way silicone vent that opens under internal pressure. (b) Flow domain for fan-driven air inlet and outlet into the sensor chamber. (c) Volume renderings of airflow at the aspirator outlet into the sensor chamber, shown from top and side views.

where $C_{d} \leq 1$ is the discharge coefficient (accounting for flow losses), and $v_{\text{eff}}$ is the effective inlet velocity. During static operation, $v_{\text{eff}} \approx v_{\text{fan}}$ . When the robot is in transit toward the VOC source at $v_{\text{transit}} \in \{0.05, 0.10\}\,\text{m/s}$ (as in the experiment), the effective inlet velocity is given in Eq. (14):

(14) \begin{equation} v_{\text{eff}} = v_{\text{fan}} + v_{\text{transit}} \end{equation}

This additive effect increases $Q_{\text{asp}}$ , enhancing VOC sampling during transit. The transport of VOCs from the odor source to the aspirator inlet is governed by both advection and diffusion. Hence, assuming a well-mixed zone near the inlet, the VOC flux into the sensor chamber is expressed in Eq. (15):

(15) \begin{equation} J_{\text{VOC}} = Q_{\text{asp}} \, C_{\text{VOC}} \end{equation}

where $ C_{\text{VOC}}$ represents the local VOC concentration at the sampling point, which is influenced by the source emission rate $ E_{\text{VOC}}$ , environmental dispersion processes, and airflow patterns. The TGS-series sensors employed in the setup (TGS sensors 2600, 2602, 2611, and 2620) respond proportionally to the sampled VOC concentration [86]; that is Eq. (16):

(16) \begin{equation} S_{\text{sensor}} \propto C_{\text{VOC, sampled}} \end{equation}

With this proportionality, we assume that adsorption and desorption kinetics are sufficiently fast relative to the sampling timescales, ensuring a near-instantaneous sensor response to changes in VOC concentration. To relieve the sensor chamber of VOCs during the recovery phase, compressed air at high pressure is passed through the soft tube opening in the sensor chamber, as described in Figures 5 and 4. The other end of the tube, connected from inside the base to the pressure source, is sealed with a $0.4$ cm check valve that prevents backflow of compressed air during the robot’s operation. The flush pressure was regulated to $10$ kPa in all experiments, flushing the air mixture out of the chamber via the forceful opening of the one-way relief vent shown in Figure 7(a).

3.5. Actuation control

The control system consists of a Faulhaber 3242X012BX4 3,692 brushless DC motor with an IER3-10000L encoder and a $20$ : $1$ gearbox paired with a Faulhaber MCBL3006 S RS driver to power the system (Figures 1 and 8). Motor actuation is managed through a $12$ -bit Dynomotion Kanalog-KFlop 8-axis controller board (Figure 8 (left)), using digital-to-analog commands. As discussed in Section 3.1, the electronic components in the base, including the pressure sensor end pins and motor driver cables, are connected to the controller with two $6$ contact airtight wire connectors. The Dynomotion board communicates with the computer via a USB Type A-Type B cable. To achieve seamless eversion without slack in the response of the spool, we calculated the required speed at a pressure of $8.8$ kPa, at which new material is added to the tip when constrained by the mount rollers. We used an acceleration-based velocity control algorithm to determine this speed. First, we derived the controller derivation by defining the error function $\epsilon _v$ between the reference velocity $\dot {q}^{ref}$ and the response velocity $\dot {q}^{res}$ as in Eq. (17):

(17) \begin{equation} \epsilon _v = \dot {q}^{ref} - \dot {q}^{res}. \end{equation}

Figure 8. (a) Dynomotion Kanalog-KFlop controller board with four motor drivers for the actuation (b) Schematics of the control unit for both the sensors and the robotic system using the Dynomotion Kanalog-KFlop controller board.

We then apply a first-order differential equation (Eq. 18) to force the error function to converge to zero and obtain the acceleration terms:

(18) \begin{equation} \dot {\epsilon }_v + k\epsilon _v = 0. \end{equation}

where $k$ is a positive coefficient to ensure that the error converges to zero. Substituting Eq. (17) into Eq. (18) yields:

(19) \begin{equation} \ddot {q}^{ref}-\ddot {q}^{res} + k(\dot {q}^{ref} - \dot {q}^{res}) = 0. \end{equation}

Since our system has a disturbance observer structure [Reference Sariyildiz and Ohnishi87], we can assume that all disturbances are compensated and that the response acceleration observed at the output of the system is the same as the input acceleration we provide to our system. Thus, the desired acceleration is obtained from Eq. (19) as follows:

(20) \begin{equation} \ddot {q}^{des}=\ddot {q}^{ref} + k(\dot {q}^{ref} - \dot {q}^{res}). \end{equation}

Eq. (20) describes the controller equation that provides the desired acceleration to the system, allowing it to follow the velocity reference. After configuring the motor speed, we applied a constant velocity reference of 0.57 $\pi$ rad/s and a voltage of $12$ V to the motor driving the main spool $M_1$ . This allowed the robot to extend up to $2$ cm in $1$ s at a pressure of $8.8$ kPa.

3.6. Sensor description, experiment setup, and exposure timeline

The TGS sensors used in this study are the Figaro MOS TGS2600, TGS2602, TGS2611, and TGS2620 [86] shown in Figure 9(a). The sensors require two separate voltage inputs: the heater voltage ( $V_h$ ) and the circuit voltage ( $V_c$ ). The heater voltage $V_h$ powers the built-in heating element, which maintains the sensing layer at an optimal operating temperature for VOC detection, while the circuit voltage $V_c$ enables measurement of the output voltage ( $V_{out}$ ) that varies with gas concentration [86]. Newly installed TGS sensors require continuous powering of the $V_h$ for at least seven days to stabilize the baseline resistance and response characteristics of the sensing element. Hence, we connected, powered, and conditioned each sensor for a full 7-day period before calibration and experimental measurements commenced.

Furthermore, the sensors are analog gas sensors that output a voltage signal corresponding to the detected gas concentration. The sensor outputs, ranging from $0$ to $5$ V, were digitized using the 12-bit analog-to-digital converter (ADC) embedded within the controller described in Section 3.5. Since the voltage output varies with gas concentration, the resulting ADC values serve as an indirect measure of the gas levels detected by the robot.

Figure 9. Experimental setup. (a) Sensor array connections and passive aspirator tubing are integrated inside the fully retracted robot body. The assembly can be mounted directly onto the tip mount in its current position or positioned prior to retraction, as illustrated on the right. (b) Complete experimental setup comprising a $100$ cm acrylic tube through which the robot everts, an odor-source platform holding four VOC containers, and a $5$ V, $5$ W fan to promote rapid volatile dispersion and improve sensor response time. (c) Static sensing: The robot advances toward the odor source at a constant speed of $2$ cm/s, stopping at distances of $20$ , $40$ , and $80$ cm for sampling. (d) Transit sensing: The robot traverses the VOC sources to a total length of $100$ cm at actuation speeds of $5$ cm/s and $10$ cm/s, with continuous sensor data acquisition during motion.

As shown in Figure 9(b), the experimental platform consists of a transparent acrylic tube with an outer diameter of $25$ cm and an inner diameter of $24$ cm, through which our $8.5$ cm diameter olfactory SGR was deployed. At the far end of the tube is a VOC source platform containing four cups, each filled with a different VOC: methane, ethanol, acetone, and gin. A portable $5$ V, $5$ W fan was positioned directly above the cups and operated at maximum speed to promote plume generation and dispersion along the tube. As discussed in Section 3.4, the robot’s tip mount contains a passive aspirator that directs incoming air containing entrained VOCs into the sealed chamber housing the four gas sensors. All experiments were carried out at room temperature ( $\approx 25^\circ$ C). Figure 10 describes the experimental protocol, with the experiment trials as follows:

Experiment 1– In-transit plus static sampling (constant speed of $2 cm/s$ ): Case 1: After the $20$ s baseline, the robot moves for $10$ s ( $20$ – $30$ s), covering $20$ cm, then stops to continue static sampling for the remaining $50$ s ( $30$ – $80$ s). Static sampling occurs $80$ cm from the odor source. Case 2: The robot moves for $20$ s ( $20$ – $40$ s, $40$ cm) before stopping, with static sampling at $60$ cm from the source. Case 3: The robot moves for $40$ s ( $20$ – $60$ s, $80$ cm), then samples statically at $20$ cm from the source. Each experiment exposes the robot to a single VOC, recording responses from all four sensors. Repeating all VOCs across three distances results in $12$ readings, each performed three times, totaling $36$ samples.

Experiment 2 – Full In-transit sampling: Here, the robot moves continuously toward the 100 cm-distant VOC at $5$ and $10$ cm/s. After a $20$ s baseline, it samples in-transit until reaching the source. Each VOC is tested separately, recording all sensor responses. Repeating for four VOCs at two speeds yields $8$ readings, each repeated three times, totaling $24$ samples.

Figure 10. Experimental timeline showing four sequential phases: Baseline (0–20 s): sensors stabilize in clean air with VOC containers sealed. Exposure (20–80 s): containers unsealed; odor sampling conducted under three in-transit + static distances – 20 cm ( $t_1 = 10$ s transit + 50 s static), 40 cm ( $t_2 = 20$  s transit + 40 s static), and 80 cm ( $t_3 = 40$ s transit + 20 s static) – and two dynamic speeds (5 cm/s and 10 cm/s over 100 cm). Residual response (80–100 s): VOC sources resealed; sensors record lingering signal decay. Recovery (100–200 s): robot returns to clean air for full signal restoration before the next trial.

5.1. Pneumatic eversion under varying configurations

The pressure dynamics of the SGR during growth were systematically evaluated under varying mechanical configurations to assess the influence of payload and actuation strategy on the system performance. As a baseline, we first measured the pneumatic requirement for a plain robot without any tip-mounted device. In this case, the pressure required for minimal growth was $2.1$ kPa, increasing steadily along the growth trajectory and stabilizing at approximately $6$ kPa for full extension to $100$ cm (Figure 14(a)). Subsequently, we evaluated three configurations with the tip-mount device: (i) all three motors ( $M_1$ , $M_2$ , and $M_3$ ) active; (ii) only the central motor $M_1$ active, with $M_2$ and $M_3$ disabled; and (iii) with embedded pouches. When all three motors were active, the initial pressure requirement was $4.2$ kPa, rising steadily to $9$ kPa at full extension. In contrast, operating only the central motor produced the steepest pressure curve, with the requirement increasing from $6$ kPa at minimal growth to approximately $14$ kPa at full length. The lack of lateral mechanical support in this configuration led to asymmetric loading, higher drag, and reduced eversion efficiency. Finally, in the embedded pouch configuration, the initial pressure was higher ( $5.8$ kPa) due to the added mass of the pouch assembly; however, the peak pressure during growth dropped to $11.5$ kPa, with the pressure curve flattening earlier in the trajectory compared to the central-motor-only scenario.

Figure 14. (a) Pressure dynamics for various growth configurations to understand how payload influences the performance of the robotic system (b) pressure needed to vertically lift the robot’s tip-mounted payload at heights of $20$ cm and $100$ cm, while comparing plain and pouch-embedded bodies.

Figure 14(b) shows the pressure needed to vertically lift the robot’s tip-mounted payload at two heights of $20$ cm and $100$ cm, comparing the two body types. In the plain body, pressure rises sharply from $4.8$ kPa at $20$ cm to $14.0$ kPa at $100$ cm, indicating an increasingly demanding lift due to cumulative gravitational torque. On the other hand, the pouch-enhanced, sectioned body offers better load support for the second section (plain), resulting in higher initial pressure ( $5.8$ kPa), due to added material weight, but lower pressure at full lift ( $8.5$ kPa) compared to the fully plain body.

5.2. Static and transit sensing

Figure 15 shows the oSGR prototype during the experiments. In both Experiments 1 and 2, all four sensors produced their strongest responses to at least one of the four target gases, while also exhibiting measurable cross-sensitivities to non-target gases. This behavior is consistent with the broad spectral width and overlapping response bands of tin-oxide–based sensors, where the observed signals reflect a mixture of selective and non-selective interactions. We explicitly consider these characteristics in the design of the classification models.

Figure 15. VOC sampling experiments. The robot grows to lengths of $20$ cm, $40$ cm, and $80$ cm toward the VOC source at a constant speed of $2$ cm/s, without contacting the pipe walls. The blue arrow beneath the tip mount indicates its height above the inner floor of the pipe. In the $80$ cm experiment (bottom), a smaller red arrow beneath the blue highlights the change in tip mount height caused by the increased pressure required to counteract the payload at the tip.

Figure 16. Experiment 1 showing sampling responses of all four sensors at three different exposure intervals: (left) $t_1$ : 10 s in-transit followed by 50 s static, (middle) $t_2$ : 20 s in-transit followed by 40 s static, and (right) $t_3$ : 40 s in-transit followed by 20 s static. Readings were recorded during exposure to methane, ethanol, gin, and acetone (a–d), respectively.

Figure 16 presents sensor responses in Experiment 1 under combined in-transit and static sampling at three distances. In case 1 (left), corresponding to short transit ( $10$ s) and long static exposure ( $50$  s), the robot began sampling VOCs after the $20$ s baseline period, advanced for $10$ s ( $20$ – $30$ s), and then continued static sampling for $50$ s. Case 2 (middle) represents moderate transit and static exposure: the robot extended $40$ cm over $20$ s ( $20$ – $40$ s) before pausing for a $40$ s static phase ( $40$ – $80$ s). Case 3 (right) depicts extended transit and short static exposure, where the robot advanced for $40$ s ( $20$ – $60$ s) and stopped at a $20$ cm standoff.

As shown in Figure 16(a–d), response magnitudes were strongly monotonic with distance. Closer standoffs consistently produced higher peak amplitudes, faster rise times, and longer durations above threshold. For instance, $80$ cm growth ( $20$ cm standoff) yielded the largest signals, whereas $20$ cm growth ( $80$ cm standoff) produced the weakest. Although decay constants were broadly similar across distances, stronger peaks resulted in proportionally larger residual signals, suggesting that higher exposures slowed complete desorption. Sensor-specific differences were also evident: ethanol and gin elicited the largest responses on TGS2602 and TGS2620, methane responses were strongest on TGS2611, and acetone produced moderate but consistent activity on TGS2602 and TGS2620.

In Experiment 2, we extended the sequence of olfactory tasks to investigate the effect of traversal speed ( $5$ cm/s, solid lines; $10$ cm/s, dashed lines) on sensor array performance. As in Experiment 1, baseline signals remained stable until exposure (Figure 17(a–d)). However, unlike static sampling, these trials exhibited compressed exposure windows due to reduced residence time in the high-concentration zone. The resulting responses were more transient, with narrower peaks and sharper decays. A clear speed-dependent effect emerged: runs at $10$ cm/s consistently showed narrower temporal profiles than $5$ cm/s, while peak amplitudes remained broadly comparable across sensor–gas pairs.

At the level of individual sensors, Experiment 2 reproduced the patterns observed in Experiment 1, although with compressed dynamics. TGS2611 again showed the strongest selectivity to methane, with distinct exposure and recovery phases that mirrored its dominant role in Experiment 1. TGS2602 and TGS2620 produced the highest responses to ethanol and gin, with peak magnitudes comparable to static cases but faster post-exposure decay. Acetone responses remained moderate on TGS2602 and TGS2620 and weak on TGS2600 and TGS2611, similar to Experiment 1. Importantly, the overall ranking of sensor–gas sensitivities was preserved across both experiments, indicating that traversal speed primarily affected the temporal shape of response curves without altering the relative selectivity of the sensor array.

Figure 17. Experiment 2: Sampling responses of the four TGS sensors at two different exposure speeds; $5$ cm/s (solid lines) and $10$ cm/s (dashed lines), during exposure to methane, ethanol, gin, and acetone (a–d), respectively. The legend colors correspond to the same sensors as shown in experiment 1.