Kiri-Capsule: Capsule Bioengineering Innovations

• Kiri-Capsule is a diverse capsule platform spanning ingestible biopsy robots, programmable microcapsules, and integrated neural inference systems.
• It employs kirigami-inspired design principles that enable controlled tissue penetration, deployable mechanisms, and safe operation under spatial and power constraints.
• The platform bridges physical, chemical, and computational domains, supporting pH-responsive encapsulation, tailored rheology, and on-device deep learning for real-time diagnostics.

Kiri-Capsule denotes, in the most specific sense, a kirigami-inspired swallowable gastrointestinal biopsy robot that couples deployable polyimide flaps, controlled shallow penetration, rotary scraping, and internal specimen retention to obtain histology-ready tissue samples during capsule endoscopy (Zhao et al., 5 Feb 2026). In a broader cross-disciplinary sense, the supplied literature supports using “Kiri-Capsule” as an Editor’s term for several capsule-centered research lines in which functionality is governed by shell architecture, hierarchical compartmentalization, deployable mechanics, or capsule-scale embedded computation. Under that broader usage, the term spans porous cellulose nanofibril microcapsules with programmable suspension rheology, brick-and-mortar colloidosomes, plant-protein hierarchical microcapsules, mechanistic models of multilayer releasing capsules, and ingestible electronic or vision-processing capsules (Dhand et al., 2021, Radulova et al., 2019, Dinh et al., 2024, Onofri et al., 20 Jun 2025, Krumb et al., 30 Apr 2025).

1. Terminology, scope, and disciplinary boundaries

The literature is heterogeneous, and the term is not used uniformly. The explicit designation Kiri-Capsule appears in the gastrointestinal biopsy robot paper, where it refers to a swallowable kirigami capsule robot for minimally invasive tissue collection (Zhao et al., 5 Feb 2026). By contrast, the cellulose nanofibril microcapsule study does not explicitly brand its system as Kiri-Capsule, although the supplied details treat that CNF/oleylamine shell concept as a natural extension of the name; likewise, the plant-protein and multi-stratum release papers do not use the term explicitly, but they describe capsule platforms that fit a broader capsule-engineering interpretation (Dhand et al., 2021, Dinh et al., 2024, Onofri et al., 20 Jun 2025).

A common misconception is that every paper containing the word “capsule” belongs to the same technical lineage. The supplied machine-learning paper on HGWCapsule does not use the name Kiri-Capsule at all; it concerns capsule learning in neural networks, where “capsule” denotes a representation and routing mechanism rather than a physical shell, ingestible device, or encapsulation carrier (&&&10&&&). The endoscopy paper eNCApsulate is closer in application domain, since it targets wireless capsule endoscopes, but its contribution is on-device bleeding segmentation and monocular depth estimation rather than capsule mechanics or biopsy hardware (Krumb et al., 30 Apr 2025).

This terminological dispersion suggests that Kiri-Capsule is best understood as a family resemblance across physical capsule technologies, not as a single standardized platform. Within that family, the strongest recurrent themes are controlled interface formation, internal-phase or shell-mediated functionality, safe deployment in constrained biological environments, and operation under severe size, power, or transport constraints.

2. Porous cellulose nanofibril microcapsules and rheological programmability

One major material interpretation of Kiri-Capsule is the cellulose nanofibril/oleylamine porous microcapsule. These capsules are liquid droplets encapsulated by a thin solid shell formed by interfacial complexation of TEMPO-modified cellulose nanofibrils with oleylamine at an oil-water interface (Dhand et al., 2021). The paper first demonstrates formation in a millifluidic T-junction using dilute CNF suspension in an OA/toluene continuous phase, with at least 5 minutes residence time for shell formation, yielding relatively uniform capsules of about $580 \pm 200\,\mu\mathrm{m}$. For rheology-scale production, it then shifts to batch high-shear emulsification at $\phi = 0.15$, mixing at 3000 or 10,000 rpm for about two minutes and then stirring for another five minutes.

The shell is both mechanically stabilizing and porous. After transfer from oil into water, FITC-dextran (10 kDa) diffuses out over time, whereas $2\,\mu\mathrm{m}$ microparticles remain trapped, indicating pore sizes on the order of 100 nm to $1\,\mu\mathrm{m}$ (Dhand et al., 2021). This combination of solid-shell integrity and selective permeability is central to the platform’s delivery and responsive-material behavior.

For suspensions of neat CNF/OA capsules, the rheology is primarily viscous and shear-thinning. The zero-shear viscosity is fitted by a modified Krieger–Dougherty relation,

$$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$

with $\phi_m = 0.73 \pm 0.22$ and $[\eta] = 4.1 \pm 1.6$ (Dhand et al., 2021). These values differ from hard-sphere reference values, and the paper attributes that deviation to softness, anisotropy, polydispersity, and possibly weak attractions. Batch-made capsules are also described as somewhat non-spherical and polydisperse, which is consistent with that interpretation.

A central result is that changing only the internal aqueous phase can transform the bulk suspension response. When polyacrylic acid (PAA) is added before capsule formation, the suspension becomes elastic, exhibits $G' > G''$, and develops a clear yield stress well described by Herschel–Bulkley behavior,

$\sigma = \sigma_y + K\dot{\gamma}^n.$

The yield stress scales as

$$\sigma_y \sim c_{\mathrm{PAA}}^{\,1.5 \pm 0.2}, \qquad \sigma_y \sim \phi^{\,2.8 \pm 0.2}.$$

At high PAA concentration, the flow curves show a viscosity plateau and hysteresis between up- and down-shear sweeps (Dhand et al., 2021).

The proposed mechanism is not depletion. Addition of polystyrene to the continuous phase produces only weak viscosity changes and no clear yield stress. Instead, the paper argues that PAA becomes incorporated into the porous CNF shell, making capsules “sticky” and inducing associative, shell-mediated capsule-capsule attractions. Under quiescent conditions, these attractions generate a percolating elastic network; under shear, the network breaks down into dispersed clusters and then individual capsules. A plausible implication is that Kiri-Capsule-type soft-particle suspensions can be programmed through cargo composition rather than external rheology modifiers.

3. Encapsulation architectures: brick-and-mortar colloidosomes, plant-protein hierarchies, and multi-stratum release models

A second major branch of Kiri-Capsule research concerns shell design for transport control. In the brick-and-mortar colloidosome system, an oil drop forms the core, silica particles form the shell “bricks,” and mixed polymer + surfactant adsorption layers form the “mortar” that blocks interparticle gaps and suppresses leakage (Radulova et al., 2019). The preparation is intentionally simple: hydrophilic silica particles (Excelica UF320, average diameter $\phi = 0.15$0) are partially hydrophobized by potassium oleate, combined with polymer and salt, loaded with oil or fragrance, and processed by a single ultrasound homogenization step. Only capsules surviving sedimentation and two water rinses are treated as stable.

The system is explicitly pH responsive. Capsules are stable in aqueous media at pH 3–6 and destabilize above pH 6, with polymer desorption, surfactant loss, particle hydrophilization, shell failure, and cargo release forming the stated mechanism (Radulova et al., 2019). Polymer choice is decisive: Carbopol 971P, Carbopol 974P, Carbopol 980, and PAA yield stable capsules, while Carbopol 971P gives the best overall robustness, with average capsule diameter about $\phi = 0.15$1 and stability for at least 8 months in water at room temperature and pH $\phi = 0.15$2. Successful encapsulation is reported for tetradecane, limonene, benzyl salicylate, citronellol, and also sunflower oil, whereas oils with either zero or higher water solubility do not yield stable capsules under the reported conditions (Radulova et al., 2019).

The hierarchical plant protein microcapsule platform extends this design logic to sustainable and biodegradable systems (Dinh et al., 2024). Using droplet microfluidics with two connected flow-focusing PDMS chips—one hydrophilic, one hydrophobic—the platform generates simple core-shell, multicore, and particle/oil/particle architectures. The shell material is based on soy protein isolate or pea protein isolate, dissolved at 10% w/v in 42% v/v aqueous acetic acid, sonicated, heated to 90°C for 40 min, and then cooled to form $\phi = 0.15$3-sheet-rich fibrillar aggregates and a hydrogel network. The architecture supports simultaneous loading of hydrophilic cargo such as vitamin C, riboflavin, vitamin B$\phi = 0.15$4, vitamin B$\phi = 0.15$5, iron, and fluorescein, together with hydrophobic cargo such as fragrances, vitamin D, vitamin E, and essential oils.

The key control variable is osmotic balance. Without CaCl$\phi = 0.15$6 in the internal phase, primary emulsions are unstable after 1 day and fluorescein is completely released over 1 month; with CaCl$\phi = 0.15$7, emulsions remain stable after 1 day and no fluorescein release is observed even after 1 month (Dinh et al., 2024). The optimized concentration is 0.5 M CaCl$\phi = 0.15$8, which also yields the best hardness and structural integrity. Digestibility by the Boisen protocol is $\phi = 0.15$9 for plant-protein microcapsules versus $2\,\mu\mathrm{m}$0 for control PPI, and biodegradability under ISO 14851 freshwater conditions reaches 98.0% relative to cellulose, exceeding the stated 90% benchmark (Dinh et al., 2024).

A complementary theoretical layer is provided by the multi-stratum spherical microcapsule model (Onofri et al., 20 Jun 2025). Here the capsule is treated as a concentric core-shell carrier with arbitrary numbers of strata, anisotropic diffusion, retention/decay, finite interfacial transfer, and surface erosion. The governing transport equation is written as

$2\,\mu\mathrm{m}$1

with radial anisotropy introduced through direction-dependent diffusivities $2\,\mu\mathrm{m}$2 and $2\,\mu\mathrm{m}$3, and anisotropy factor

$2\,\mu\mathrm{m}$4

In the reported alginate-based case study, the capsule radius is about $2\,\mu\mathrm{m}$5 and shell strata are nanometric, around 18 nm (Onofri et al., 20 Jun 2025). Validation against INFOGEST 2.0 digestion experiments reproduces negligible oral-phase release, about 21.92% release in the gastric phase, and sustained intestinal release, with fitted parameters including $2\,\mu\mathrm{m}$6 and early-layer anisotropy $2\,\mu\mathrm{m}$7. This suggests that staged diffusivity and erosion are not merely descriptive variables but active design parameters for Kiri-Capsule-type release systems.

4. Capsule endoscopy and on-capsule neural inference

In wireless capsule endoscopy, Kiri-Capsule can also denote capsule-scale computational autonomy rather than shell chemistry. The eNCApsulate framework addresses two practical bottlenecks in WCE: large video volume and poor localization after ingestion (Krumb et al., 30 Apr 2025). Capsules record hours of video, physicians may spend 30–120 minutes reviewing a single study, and existing localization often requires extra internal sensors or external magnetic arrays. The paper therefore aims to move both pathology detection and localization-related depth estimation directly onto the capsule.

The computational core is Neural Cellular Automata (NCA). Each cell observes its Moore neighborhood via $2\,\mu\mathrm{m}$8 filters; neighborhood features are concatenated and processed by an MLP; the MLP predicts an update vector; the update is added back into the image/state buffer; and only a stochastic subset of cells is updated at each step (Krumb et al., 30 Apr 2025). The two reported variants are eNCApsulateS for bleeding segmentation and eNCApsulateD for monocular depth estimation, using 18 and 22 total channels respectively. RGB occupies the first three channels, hidden channels are in the middle, and an output channel is placed at the end. Hidden channels are initialized with noise, which the authors state improves robustness.

Training is performed in PyTorch on a PC with an NVIDIA GeForce GTX 3090. Segmentation uses the KID2 dataset with supervised labels. Depth estimation uses distillation from Depth Anything V2 on a small subset of KID2: 430 depth maps judged too flat are removed, using a normalized gradient magnitude threshold that accepts maps only if the gradient magnitude exceeds 1.1, leaving 727 annotated samples for training (Krumb et al., 30 Apr 2025). The depth loss combines MSE, SSIM, and image-gradient terms with weights

$2\,\mu\mathrm{m}$9

Minibatches start at size 8 and are duplicated to 16; random crops begin at $1\,\mu\mathrm{m}$0 and are downsampled to $1\,\mu\mathrm{m}$1.

The deployment target is an ESP32-S3 microcontroller. Inference is reimplemented in ANSI C, exploiting SIMD instructions and the FPU. Runtime optimizations include executing the stochastic update first, avoiding separate intermediate convolution buffers, and using only two buffers: an image buffer and an update buffer (Krumb et al., 30 Apr 2025). On the reported benchmarks, eNCApsulateS achieves Dice $1\,\mu\mathrm{m}$2 and IoU $1\,\mu\mathrm{m}$3, compared with the best lightweight baseline Dice of 0.687. The model size is reported as 44.32 kB in one table representation and 47,152 bytes in another, while the paper states that NCA requires more than 100× fewer parameters stored in memory than other small-scale models (Krumb et al., 30 Apr 2025).

On-device performance is central. Average inference speed on ESP32-S3 improves from 9 s to 3 s per image with SIMD optimizations, and with temporal regularization it decreases from 3 s to less than 1 s per image on average (Krumb et al., 30 Apr 2025). Each NCA time step is roughly 65 ms, and a typical NCA may need around 100 steps to converge without optimization. Temporal regularization applies early stopping after at least 10 steps when hidden-state change falls below 0.1, with a cooldown counter of 5; on a bleeding video, this reduces total NCA steps from 6,988,560 to 1,222,998 while preserving segmentation quality. The paper presents this as the first work enabling reliable bleeding segmentation and depth estimation on a miniaturized capsule-class device itself.

5. Kiri-Capsule as a kirigami-inspired gastrointestinal biopsy robot

The most explicit and mechanically distinctive use of the term is the Bioinspired Kirigami Capsule Robot for Minimally Invasive Gastrointestinal Biopsy (Zhao et al., 5 Feb 2026). This capsule is designed to close the diagnostic gap between passive imaging and histology-ready tissue acquisition in WCE. Its reported dimensions are approximately 17 mm in diameter and 22 mm in length, compatible with ingestible devices. The architecture includes a front-end biopsy actuator driven by a miniature four-phase stepper motor, cam 1 and cam 2, a rotating shell, rotating shaft, blade, connecting rod, and protective shells. The biopsy skin is laser-cut from polyimide (PI) film.

The kirigami pattern is arranged on a triangular lattice with parameters $1\,\mu\mathrm{m}$4 mm, $1\,\mu\mathrm{m}$5 mm, and $1\,\mu\mathrm{m}$6. Strip dimensions are $1\,\mu\mathrm{m}$7 mm and $1\,\mu\mathrm{m}$8 mm, and four PI thicknesses are tested: $1\,\mu\mathrm{m}$9 mm (Zhao et al., 5 Feb 2026). During locomotion the PI sheet remains essentially flat, preserving compactness. Under actuation, a dual-cam mechanism stretches the sheet, transforming in-plane cuts into sharp out-of-plane protrusions. Continued rotation then produces rotary scraping. Positive motor polarity causes deployment followed by scraping, while reverse polarity retracts the flaps.

The cam-driven kinematics are modeled explicitly. The pressure angle is

$$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$0

and the follower kinematics are

$$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$1

These relations are used to justify bounded pressure angles and controlled acceleration during deployment and retraction (Zhao et al., 5 Feb 2026).

Material characterization shows that pristine PI film has a Young’s modulus of approximately 20 MPa. Thickness strongly affects actuation: $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$2 and $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$3 mm are too stiff for capsule-scale morphing, whereas $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$4 and $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$5 mm undergo smoother transitions; the $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$6 mm film is selected for subsequent experiments (Zhao et al., 5 Feb 2026). Deployment is quantified by the flap opening angle $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$7 as a function of strain $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$8: at $$\frac{\eta_0}{\eta_s} = \left(1-\frac{\phi}{\phi_m}\right)^{-[\eta]\phi_m},$$9, flaps reach approximately $\phi_m = 0.73 \pm 0.22$0; at $\phi_m = 0.73 \pm 0.22$1, about $\phi_m = 0.73 \pm 0.22$2.

Penetration is intentionally shallow. The theoretical penetration depth is $\phi_m = 0.73 \pm 0.22$3 mm based on $\phi_m = 0.73 \pm 0.22$4. In ex vivo porcine tissue experiments $\phi_m = 0.73 \pm 0.22$5, the median measured penetration depth is approximately 0.61 mm, with interquartile range 0.51–0.65 mm, range 0.46–0.66 mm, and mean 0.57 mm (Zhao et al., 5 Feb 2026). Force measurements show that against porcine gastric tissue, the $\phi_m = 0.73 \pm 0.22$6 and $\phi_m = 0.73 \pm 0.22$7 components increase to about 0.5–2.0 N over a 2 s interval; for small intestine, peaks are about 0.3–1.0 N around 1.5 s. These values are stated to remain within safe or typical GI biopsy thresholds.

Biopsy performance is reported from seven biopsies per tissue type. Gastric samples have median mass 10.7 mg, IQR 9.4–12.6 mg, range 6.2–15.1 mg, and mean approximately 10.9 mg. Small-intestinal samples have median mass 18.2 mg, IQR 17.0–20.8 mg, range 14.4–23.9 mg, and mean approximately 18.9 mg (Zhao et al., 5 Feb 2026). Hematoxylin and eosin sections show that both mucosa and submucosa are captured. The retained tissue is stored in internal fan-shaped cavities, which act as specimen reservoirs. The main limitation identified in the supplied details is that the current prototype is tethered and requires wired power, with future directions including untethered operation, potentially via magnetic actuation, and multi-segment designs for multi-site sampling.

6. Kirigami-structured electronic capsules for long-term gastric residence and monitoring

A further development of the Kiri-Capsule concept is the kirigami-structured electronic capsule for long-term gastric monitoring (Huang et al., 7 May 2026). This is a gastric-resident ingestible robotic platform designed for week-long operation, though the supplied swine details report stable residence confirmed by X-ray on day 0, day 9, day 20, and day 30. The architecture separates a rigid central domain, containing the battery pack, power management, and wireless modules, from a deployable domain, containing three superelastic nitinol arms, a kirigami-patterned flexible PCB, distributed sensors, and an electrically triggered release mechanism.

The kirigami implementation is a flexible printed circuit board spanning the capsule body and the arms. Distributed cuts localize strain away from copper traces, permit out-of-plane deformation, reduce the bending radius, and preserve electrical continuity during arm rotation and gastric deformation (Huang et al., 7 May 2026). The paper reports that the kirigami design tolerates approximately $\phi_m = 0.73 \pm 0.22$8 arm rotation, corresponding to roughly 2 mm extension. Among the tested 3-cut, 4-cut, and 5-cut geometries, the 4-cut pattern is selected as the best compromise: the 3-cut design only marginally meets the target and fractures at about 3 mm extension, whereas 4-cut and 5-cut survive but the 5-cut consumes more layout area.

Retention and release are governed mechanically by deployable nitinol arms and a polycaprolactone (PCL) locking rod. In a funnel-based radial compression test, it requires more than 5 N to push the deployed device through the funnel, exceeding the reported physiological radial compressive forces of 0.3 N to 1.6 N in the stomach/pylorus (Huang et al., 7 May 2026). The PCL rod has diameter $\phi_m = 0.73 \pm 0.22$9 and tensile failure load approximately 25 N, providing the stated safety margin. Release is chemistry-independent and instead uses Joule heating to soften the thermally responsive PCL. At 100 mA, total release energy is reduced to about 100 mJ; a practical operating point of 60 mA is chosen to balance reliability and battery lifetime. Total release energy is approximately 0.1 to 0.3 J, compared with about 800 J stored in a representative 3.7 V, 60 mAh battery pack (Huang et al., 7 May 2026).

The demonstrated application is continuous gastric radiation dosimetry. The capsule integrates a photodiode array for radiation sensing, an IMU for motility monitoring, a clinical-grade temperature sensor for calibration, and a metal oxide gas sensor (Huang et al., 7 May 2026). Two photodiode readout modes are reported: direct ADC sampling and a threshold-triggered amplified circuit (TTAC). Projected battery life is 98 hours for the ADC architecture and approximately 2850 hours for the interrupt-driven amplified architecture, a roughly 30-fold improvement.

Wireless communication is treated as a first-class systems problem because the stomach is electromagnetically lossy and dynamically variable. The platform uses dual-band Bluetooth Low Energy at 2.4 GHz and 915 MHz sub-GHz RF, with RSSI- and throughput-aware adaptive transmission implemented as a dual-loop PID controller (Huang et al., 7 May 2026). In water as a tissue surrogate, BLE RSSI drops from about $[\eta] = 4.1 \pm 1.6$0 in air reference to about $[\eta] = 4.1 \pm 1.6$1 at 2 cm immersion and approaches $[\eta] = 4.1 \pm 1.6$2 at 6 cm immersion, whereas 915 MHz stays around $[\eta] = 4.1 \pm 1.6$3 at 2 cm and remains above $[\eta] = 4.1 \pm 1.6$4 at 6 cm. With adaptive control, RSSI is maintained near $[\eta] = 4.1 \pm 1.6$5, throughput at $[\eta] = 4.1 \pm 1.6$6, and power at $[\eta] = 4.1 \pm 1.6$7, compared with less stable fixed-power operation (Huang et al., 7 May 2026).

The reported in vivo outcomes in Bama miniature pigs include stable gastric residence, real-time telemetry over seven days within roughly $[\eta] = 4.1 \pm 1.6$8 to $[\eta] = 4.1 \pm 1.6$9 without progressive degradation, repeated radiation-induced peaks on days 1–4, and successful triggered disassembly (Huang et al., 7 May 2026). After release, imaging shows immediate fragmentation on day 0, separation of PCB components by day 3, entry into the small intestine by day 5, and entry into the colon by day 10. The supplied text does not describe histology for this study; safety conclusions are based on stable residence, absence of migration during retention, and unobstructed passage after disassembly.

Taken together, these results indicate that Kiri-Capsule has evolved into a broad research motif spanning porous and rheologically programmable microcapsules, pH-responsive and biodegradable encapsulation systems, multilayer release modeling, capsule-scale neural inference for endoscopy, minimally invasive kirigami biopsy robotics, and long-term gastric-resident electronic monitoring. The unifying principle is not a single material or mechanism, but the systematic use of capsule form factors to couple constrained geometry with highly engineered interfaces, deployable structures, or embedded function (Dhand et al., 2021, Radulova et al., 2019, Dinh et al., 2024, Onofri et al., 20 Jun 2025, Krumb et al., 30 Apr 2025, Zhao et al., 5 Feb 2026, Huang et al., 7 May 2026).