Intelligent magnetic manipulation for gastrointestinal ultrasound

To show feasibility of the proposed approach, we leveraged our previous work on intelligent magnetic actuation for painless colonoscopy (53) to develop a previously unexplored RCE and autonomous μUS robot control strategy. Our system (Fig. 1) consists of an extracorporeal permanent magnet (EPM) and an intracorporeal permanent magnet (IPM), the latter being embedded in the RCE. The RCE has magnetic field sensing capability that facilitates real-time localization of the device, the accuracy of which is about 2 mm and 3° (53). Throughout this manuscript, we assume this low-pass filtered localization feedback to be a ground truth. We use pose feedback for closed-loop magnetic control, where positions, orientations, and velocities of the RCE can be commanded. We note that the pose repeatability of the serial arm, and thus the EPM, is ±0.15 mm (LBR Med 14 R820, KUKA). This control resolution is thus the minimum bound on the control repeatability of the RCE. Environmental factors and field modeling have further effects on RCE repeatability. The forces and torques applied on the RCE are estimated using the point-dipole magnetic model. Using an on-board transducer and an oscilloscope, our system receives signal feedback in the form of waveforms. The logarithmic representation of these waveforms results in an A-scan. The waveforms are processed to develop an echo signal rating, or ESR. The ESR is an amplitude-based metric that determines the maximum magnitude of the received acoustic signal. It is computed by applying digital filters to a raw voltage output in digital form from μUS transducers. A high ESR for an acoustic waveform implies that high-amplitude echoes are present in the signal, and thus acoustic coupling and adequate alignment with the environment exist. The diagnostic task of determining the content of echoes is left to the operator. Our system can capture diagnostic images while the RCE moves and can use information from the A-scans to influence magnetic control. We use this feedback to autonomously capture μUS signals in an uncertain and unconstrained environment, where acoustic coupling cannot easily be confirmed. The acquisition of sequential A-scans facilitates the generation of B-scans, i.e., 2D US images with a logarithmic representation of echo amplitude. A graphical representation of our RCE moving through the colon lumen while building a B-scan is shown in Fig. 1A, and an overview of our system is shown in Fig. 1B. A video describing the overall concept can be found in the Supplementary Materials (movie S1).

We conducted a series of benchtop experiments to investigate the feasibility of this concept. It included (i) RCE control development, where tests were performed in a silicone phantom to better understand how to manipulate the RCE to acquire μUS images and characterize the relationship between control parameters and image quality; (ii) servoing using US feedback, where the feasibility of closing the robotic control loop using US echoes was explored; (iii) benchtop system validation, where the ESR method was validated on an acoustically realistic agar phantom (the demonstration of the system’s ability to autonomously acquire spatially relevant US images was also demonstrated during these tests); and (iv) in vivo validation and demonstration of an RCE acquiring μUS waveforms in a living porcine model.

We first characterized the robotic system by studying the relationship between RCE control parameters—i.e., tilt, roll, and RCE-substrate magnetic contact force—and μUS images. We found that the successful acquisition of μUS signals of sufficient amplitude by our RCE is sensitive to four main criteria—all with respect to the substrate: (i) the presence of an US coupling medium, (ii) transducer-tissue contact force, (iii) transducer tilt, and (iv) transducer roll. The conventions used for the tilt and roll of the RCE can be seen in fig. S1A. We conducted benchtop experiments using a silicone phantom (fig. S1B) to characterize this sensitivity and validate the system’s ability to effectively manipulate the transducers (i.e., RCE).

We conducted 10 instances of each test, with an operator teleoperating the EPM (i.e., serial manipulator): increasing coupling force from 0.25 to 2.5 N (RCE tilt = 0°), adjusting RCE tilt from −10° to 10° (coupling force = 0.6 N), and adjusting RCE roll in a span of 50° (coupling force = 0.6 N). We found that a higher force can strengthen US signals; however, the dependence of signals on coupling force is not linear or repeatable (Fig. 2A). We found that strong acoustic coupling exists in a ±3° tilt range of the RCE. In roll characterization trials, we found that acoustic coupling was strong in a roll range of about ±10.5°. These results show that contact force facilitates acoustic coupling (tight contact), and tilt and roll determine the transducer alignment, which is necessary to ensure that reflected echoes are successfully received.

Sample-filtered time series of waveforms used for characterizing the force, tilt, and roll ranges are shown in Fig. 2A and figs. S2 (A and B) and S3A. The signal strength of μUS waveforms is subject to considerable attenuation and scattering without a coupling medium (in our case, we used water and US gel). For this reason, the RCE was designed with an embedded water channel for on-demand, in situ irrigation. When considering just the substrate, acoustic coupling is likely to be better in an in vivo setting compared with our silicone phantom, owing to the continuous secretion of mucus (56). A minimum contact force should be maintained, but the force characterization results suggest that attempting to characterize a force–echo strength relationship is likely not practical owing to a lack of robust repeatability; i.e., a known force does not necessarily induce a standard echo quality. Tissue contact force should also be kept to a minimum to avoid mucosal damage. A previous study has shown that no damage was inflicted on porcine mucosa when irrigation pressures of up to 3 bar were applied (57). The maximum (worst-case) pressure expected from our RCE is 0.75 bar. This was calculated by assuming an RCE tilt angle of 45°, a tissue indentation of 1 mm, and a maximum contact force achievable by our current system of 2.5 N. During the tilt and roll characterization tests and all other experimental work, the contact force did not exceed 1.5 N.

Experiments were subsequently performed to determine whether giving calibration-based pose commands and autonomously approaching them are sufficient for robust GI μUS imaging. We used the characterization results to determine the parameters that should be controlled and appropriate ranges for them. Tests were also performed to assess whether μUS signals could be successfully acquired while the RCE moves along a linear trajectory; these were conducted using both teleoperation and autonomous navigation. Table 1 summarizes this work.

Although the RCE was able to reach desired configurations and traverse paths, reliable imaging was not always achieved. This was attributed primarily to the lack of μUS feedback and the necessity for both acoustic coupling and transducer alignment.

The acquisition of μUS echoes is dependent on successful acoustic coupling and alignment of the transducer with a substrate. Simply manipulating the RCE to a configuration or through a trajectory is insufficient for reliable μUS imaging (Table 1). Therefore, the ESR was used to provide a quantitative real-time measurement of the acoustic coupling; without this metric, a control loop could not be closed with μUS.

Our approach focuses on the amplitude of the received signal and compares it with an empirically set threshold (ESR_thresh) at and above which echoes can be clearly seen. Although an ESR below ESR_thresh may show potentially useful, but less clear, echoes, the threshold serves as a measure of the signal strength that the system strives to attain. We developed an autonomous echo-finding algorithm that uses this ESR for attaining clear acoustic images. The rating method and algorithm were evaluated during benchtop experiments. The user and autonomous system were first tasked with using the ESR to manipulate the RCE, adjusting tilt and force, to acquire echoes—i.e., echo detection—before the system’s ability to autonomously react to external disturbances—i.e., echo maintenance—was evaluated. Table 2 summarizes these tests and results. As seen in Table 2, the results suggest feasibility of our ESR method and its effectiveness as feedback for robust diagnostic imaging.

The previous experimental trials focused on the development of robotic control and the method of closing the loop with μUS feedback. Further tests were performed to assess the performance of the system on a more acoustically realistic phantom (fig. S1C): an agar-based material with embedded features. Additional information on this phantom can be found in the Supplementary Materials. These tests demonstrate the capability to find echoes autonomously using the ESR method and to combine RCE localization and manipulation to complete an autonomous linear trajectory while gathering images with clear features. The latter demonstrates the acquisition of spatially relevant US images—i.e., B-scans with a spatially accurate horizontal axis. The results from these tests are summarized in Table 3.

The echo detection results provide further support for the ESR method used. The teleoperated results were weaker than those seen on the silicone phantom, with the operator repeatedly failing to achieve reliable acoustic coupling. This was attributed to (i) the increased challenge of manipulating the capsule on a flatter, lower-friction substrate, where the curvature of the environment does not aid alignment, and (ii) the time limit set on the experiments, showing the increased cognitive burden of teleoperating the RCE (this requires the simultaneous monitoring and control of capsule pose, contact force, and ESR output with sufficient reaction speed). The spatial scanning results (with an example B-scan shown in fig. S4) suggest that it is possible to autonomously acquire spatially relevant US images by combining locomotion with autonomous echo detection—a capability that is necessary for future diagnostic tasks.

The robotic system was evaluated in an in vivo porcine model (Fig. 3A), with the primary objectives of (i) successfully merging μUS, robotic, and magnetic technologies in vivo; (ii) comparing teleoperated and autonomous probe manipulations in terms of μUS image acquisition performance; and (iii) assessing the feasibility of performing targeted μUS for GI screening in a living organism. The experiments, summarized in Table 4, included both teleoperated and autonomous echo detection experiments, where the goal was to manipulate the RCE to acquire distinct in vivo μUS images of the bowel wall.

The level of electrical interference between the subsystems was acceptable, and the RCE integrity was maintained throughout the tests. Tissue from the animals was inspected postmortem, and no visual signs of trauma were seen. These results show that teleoperation was inadequate in this context; it was associated with high cognitive burden because of the required controlling of multiple variables simultaneously and in a complex, dynamic environment. The poor teleoperated results (Table 4) also show the importance of proper transducer alignment, because inadequate alignment resulted in poor acoustic coupling and fewer received echoes, despite there being continuous magnetic attraction (RCE-tissue contact). During autonomous operation, the robot was able to find an echo and react to a change in the ESR to briefly maintain the configuration at which the echo was acquired (Fig. 3B). A video summarizing the in vivo work can be found in the Supplementary Materials (movie S3).

The system (Fig. 4) consists of a custom-built RCE with embedded μUS transducers, a serial manipulator with a permanent magnet at its end effector for applying fields and gradients to the RCE, custom circuitry for magnetic localization and force and torque estimation, a μUS data acquisition system, and a custom software system on a desktop computer. Movie S1 provides an overview of the RCE concept.

The RCE was developed and fabricated for in vivo use. It comprises a 3D-printed shell (Standard Clear resin, Form 2 3D printer, Formlabs), a camera (CMOS, MO-T1003-65-N01, Misumi), a light-emitting diode (LED) light source, an irrigation channel, a permanent magnet [cylindrical, axially magnetized, 11.11 mm in diameter and in length, neodymium iron boron (NdFeB), N52 grade, D77-N52, K&J Magnetics, USA] encased with localization circuitry [custom flex circuit (53)], and two miniature μUS transducers. The RCE was assembled by hand, sealed with epoxy (Permabond 4UV80HV), and coated with Parylene C for biocompatibility and lubricity (64). The current prototype, measuring 21 mm in diameter and 39 mm in length, includes a tether (6 mm in diameter, PEBAX material) for power and data transmission and as a fail-safe method to retrieve the device. The RCE is manipulated via the magnetic wrench between the small IPM and the larger EPM (cylindrical, axially magnetized, 101.6 mm in diameter and in length, NdFeB, N52 grade, ND_N-10195, Magnetworld AG, Germany) positioned at the end effector of a medical-grade, 7-DoF serial robotic manipulator (14-kg payload, LBR Med, KUKA).

The force estimation strategy relies on the recovery of RCE pose information. This is done using a localization strategy that generates 6–DOF, low-pass filtered, pose information at 100 Hz with about 2 mm and 3° accuracy (53). Throughout this manuscript, we assume that our localization reading is the ground truth. This assumption is justified by the validation of the method described elsewhere (53). The magnetic control method used in this work relied on linearizing magnetic wrenches and using them to compute necessary joint values of a serial manipulator; it is based on previous control work (53, 65, 66). Autonomous control routines in this study used a nonholonomic velocity controller for translation, where velocity was commanded along a trajectory and motion orthogonal to the path was treated as a position error. Orientation control was position level; i.e., specific angle tilts and pans were commanded as opposed to commanding angular velocities.

The control system was implemented using Robotic Operating System (ROS) (67). An overview of magnetic actuation is included in the Supplementary Materials. Vectors are described using bold lettering, v, and unit vectors are described using $\stackrel{̂}{\mathit{v}}$. All points and vectors are represented in the global frame. The dipole model is used for magnetic control, and the underlying equations for the magnetic field, force (f_m), and torque (τ_m), are described in the Supplementary Materials. Magnetic actuation of devices is achieved by modifying the direction and the gradient of an external field; these modifications result in changes in applied torque and force, respectively. This applied torque and force, or wrench, results in motion of the actuated device. The governing control relation for robot actuation is shown in Eq. 1; this expression is similar to previous works in the field (53, 68). This relationship is a linearization of the nonlinear relationship between magnetic wrench and EPM twist. The position and heading of the EPM are indicated using p_e and ${\stackrel{̂}{\mathit{m}}}_e$, respectively, where m_e is the magnetic moment of the magnet. The implementation of proportional-integral-derivative control is denoted by pid(). The heading of the RCE is denoted by ${\stackrel{̂}{\mathit{m}}}_i$, and the use of the subscript “d” indicates a desired value that is commanded by a higher-level control. The heading error e_h is computed using ${\stackrel{̂}{\mathit{m}}}_i\times {\stackrel{̂}{\mathit{m}}}_{i_d}$, and the error magnitude is the angle between these vectors. The magnetic actuation Jacobian is J_e ∈ ℝ^6×6. The pseudo-inverse of J_e is given by $J_e^{+}$

The desired EPM motion is then converted to joint-level commands for the serial robot using the relationship between EPM motion and joint commands [Eq. 2 (68)], where $\mathbb{S}$() indicates the skew-symmetric form of the cross-product operation, the purpose of which is to compensate for the symmetry of the EPM’s field about its magnetization axis. The identity matrix is represented as ${\mathbb{I}}_3$, and a matrix of zeroes is represented by ${\mathbb{O}}_3$. The geometric Jacobian J_r is that of the serial manipulator

The changes in joint values are computed via Eq. 3, where ξ is a small damping scalar, integrated, and sent to the serial manipulator’s internal controller

The tilt (α) and roll (γ) of the RCE are visualized in fig. S1A. The tilt is computed from Eq. 4. The roll is defined as rotation of the RCE about its major axis from a nominal orientation

We used two 5-mm-diameter PVDF (polyvinylidene fluoride) μUS transducers with a center frequency of 30 MHz and a physical focus at 6 mm (f# = 1.2) mounted on the top surface of the RCE with the faces embedded about 1.5 mm into the RCE to bring the focal point closer to the luminal surface of the bowel wall. PVDF was used instead of the more common piezoceramic because of its flexibility and transducer manufacturability; a 30-MHz transducer required a conductive film of 9 μm in thickness, making it easy to achieve a spherical focus in a small package. Data were acquired from one transducer, and a second one was installed for redundancy. The transducers were connected to a custom printed circuit board (PCB) inside the RCE. This PCB was connected to external hardware via individually shielded micro-coaxial cables (42 AWG core, 9442 WH033, Alpha Wire, USA) that provided some electrical shielding from the other RCE circuitry. The US transducers were pulsed during transmission, and the received echoes were amplified by a commercial pulser-receiver (DPR300, JSR Ultrasonics, Imaginant Inc., USA) with 70 dB gain before digitization with a two-channel, 70-MHz oscilloscope (DSOX2002A, Keysight Technologies). After digitization, signals were envelope-detected using a Hilbert transform and logarithmically compressed.

We implemented a waveform processing algorithm in Python to compute the ESR. This ROS node interfaced with an oscilloscope using the Python Universal Serial Bus Test, the Measurement Class library, and a Keysight Technologies application programming interface (API) to acquire waveforms from the oscilloscope at a maximum rate of 3 Hz. Waveform processing consisted of (i) trimming the raw waveform (an array of voltages) to a desired starting processing depth, (ii) applying a forward-backward linear filter using SciPy, (iii) flattening the signal using a low-pass filter to make relative echo magnitudes appropriate, and (iv) computing the maximum peak-to-peak waveform amplitude range in the waveform, which is the ESR. This ESR is then filtered over time. The ESR is an amplitude-based metric that informs an operator, or computing system, of the presence of acoustic information in a signal. We note that the clinical interpretation of this signal is left to the operator. Because the metric is amplitude-based, the ESR is not highly sensitive to filter parameters. To demonstrate what the μUS images that are used to compute the ESR look like, we refer the reader to figs. S3 and S5B, which are time-series images of the filtered waveforms that correspond to those shown in Figs. 2 and 3B. Here, we emphasize the relative clarity with which an echo can be identified (when comparing Fig. 2 with fig. S3). Because the μUS transducer is used as both transmitter and receiver in the system, residual energy from the transmit pulse is seen in the form of “ringdown.” We therefore implemented an additional ringdown-compensation filter by subtracting a moving average of the ringdown portion of the signal at each time step as follows

Here, ${\stackrel{\rightharpoonup }{p}}_B\in {\mathrm{ℝ}}^{\mathrm{m}}$ is a vector that represents the waveform values (voltages) of length m. The number of waveforms that are used in the filter is indicated by n. The length of this vector corresponds to the depth of an A-scan. The ringdown-filtered waveform segment is referred to by ${\stackrel{\rightharpoonup }{p}}_{\mathit{Br}}\in {\mathrm{ℝ}}^{\mathrm{m}}$. This filter assists with visualizing areas of interest at smaller tissue depths, which may be otherwise obscured by the ringdown feature. It is demonstrated in fig. S5, where fig. S5A shows the B-scan without the ringdown filter, and fig. S5B shows a time series of filtered waveforms combined with ringdown compensation. The waveforms in fig. S5B are used by our real-time algorithm to compute the ESR.

The echo-finding algorithm was implemented using a low US acquisition rate and aimed to show the feasibility of autonomous echo acquisition and optimization. Our autonomous method, illustrated in Fig. 5, continuously increments the desired tilt of the RCE while maintaining a constant, predefined, tissue contact force. The algorithm is activated while the RCE is in an arbitrary pose—with or without transducer-tissue coupling. The algorithm then goes to and maintains a force while tilting the RCE between hard-coded tilt limits. The control process obtains ESR updates from a separate echo processing script and continuously checks whether the echo is above a desired threshold (ESR_thresh). The ESR_thresh is determined before algorithm usage via a calibration procedure where an operator teleoperates the RCE until acoustic coupling is obtained. The value of the ESR that results in an adequate acoustic signal is then used as ESR_thresh. This approach was used because of the high variability between acoustic characteristics of different substrates. Once the control process observes an ESR value that is above ESR_thresh, the controller switches states from “searching” to “maintaining RCE configuration.” Here, the EPM makes motions that maintain the RCE in the pose in which distinct echoes were observed. In the case that the control process observes a loss of echo, it switches back to the “search” state (tilting routine).

Benchtop testing was performed with the primary goals of characterizing and refining the system in a known environment. For characterization trials, a phantom tissue substrate, 200 mm in length and ~7 mm in thickness, was made from cast silicone (Ecoflex 00-30 silicone and 10% Slacker by mass). The phantom was placed in a semicylindrical acrylic tube (77 mm internal diameter) and attached to an adjustable frame. The choice of silicone and its thickness were made to allow the RCE to deform into the substrate while also allowing repeatable/comparable measurements. For validation tests, an agar-based formulation was used to develop a phantom that was acoustically closer to tissue (69–71). It included fabric mesh to simulate features of convoluted shape, microscale particles to provide attenuation and speckle, and several regularly sized and spaced features along its length. A full description of this phantom can be found in the Supplementary Materials. Acoustic coupling gel was used as both a coupling medium and a lubricant. The individual performing the experiments had experience with operating the robotic arm and the construction of the RCE. The benchtop platform and a summary of the testing can be seen in movie S2.

Characterizing the effect RCE motion and force had on echo quality was done using the semicylindrical silicone substrate placed in a horizontal configuration. Given a known substrate orientation, RCE pose feedback provided the relative tilt between the transducers and the phantom. To ensure that echoes are most effectively reflected to the transducer, the transducer’s focus should be as close to normal to the substrate. Knowledge of transducer-tissue contact force, transducer tilt, and transducer roll sensitivity provides perspective as to which parameters of the RCE are important in acquiring and improving localized μUS images. Given that the magnetization axis of the magnet in the RCE is along its body, the RCE’s roll cannot be controlled. However, we placed the magnet off the center body axis of the RCE to induce a restorative torque to keep the transducers biased toward a vertical orientation (72).

The effect that force (RCE-substrate contact force) had on the μUS signal was characterized by conducting 10 force-varying motions while data were acquired from a μUS transducer; the force was increased from 0.25 to 2.5 N under teleoperation while tilt was kept as horizontal (i.e., 0°) as possible. To characterize the effect that RCE tilt had on the μUS signal, we used a similar protocol, but this time, the operator adjusted the RCE tilt between ±10° via teleoperation, whereas force was kept constant at 0.6 N. The roll characterization varied in protocol in that the RCE was rolled by manually applying a torque to the device. The roll angle was varied in a 50° range while force was kept nearly constant at 0.6 N.

Appropriate control strategies and parameter (force and tilt) values for acquiring μUS signals during RCE motion were chosen based on the results from the characterization work. In the first set of experiments, the RCE autonomously approached two configurations of force and tilt. The two configurations were chosen to show control repeatability and were (i) –2° tilt and 1-N coupling force and (ii) 3° tilt and 0.9-N coupling force. The phantom substrate was oriented horizontally. Before each repetition, the user placed the RCE into an arbitrary decoupled pose and force. The test was then initiated, and the system autonomously approached the desired configuration while μUS signals were continuously acquired. This autonomous maneuver can be seen in movie S2. Tests were done to assess the feasibility of moving the RCE through a linear trajectory while continuously acquiring localized μUS signals. The RCE was placed into a desired configuration (0° to 1° tilt and 0.6-N force) before being moved along an 8-cm linear trajectory under both teleoperated and autonomous control regimes. Both the system and user were tasked with maintaining the desired configuration during locomotion in an attempt to maintain μUS coupling without ESR feedback. Localization data could be combined with continuously acquired μUS signals to generate spatially relevant μUS images, i.e., 2D B-scans of depth and width/distance. Although our recorded position is 3D, the scans are plotted in 2D with respect to the distance traveled by the RCE.

An objective assessment of ESR was not conducted because a universal standard indicator for US image quality does not exist. The ESR was first assessed subjectively by visually confirming that increases in ESR correspond to improvements in echo amplitude seen in processed B-scans. Two test protocols were then used to assess whether the ESR could be effectively used as a form of feedback in the robotic control loop (these preliminary tests were done using the silicone phantom). In the first protocol, we conducted echo detection trials in both a horizontal phantom and a tilted phantom to simulate the unknown environment configuration that would be present during use in vivo. Both teleoperation and autonomous control were used. The RCE was first placed in an arbitrary start pose (transducers decoupled from substrate) before the tilt was adjusted within hard limits until ESR_thresh was exceeded (i.e., μUS coupling achieved). A trial was considered successful if multiple distinct μUS echoes were captured. During the autonomous tests, once the RCE found acoustic coupling, it used magnetic closed-loop control to maintain its configuration; simultaneously, the system continued to check for a loss of acoustic coupling. The second protocol was conducted to observe whether the system could autonomously react to external disturbances and attempt to maintain μUS coupling. From an arbitrary start pose, the RCE was tasked with autonomously detecting an echo and then maintaining that configuration in the presence of disturbances. Disturbances were induced either by teleoperating the EPM to override the autonomous routine and cause unwanted motion in the RCE or by pulling on the tether of the RCE to decouple it acoustically from the substrate. Movie S2 shows this use of the ESR and an example of the disturbances used.

The acoustically relevant agar phantom was used first to validate autonomous echo detection (i.e., the ESR methodology) and, second, for additional μUS servoing experiments that demonstrated combined echo detection and autonomous locomotion. In the former, we repeated the echo detection tests performed on the silicone phantom under both teleoperated and autonomous control regimes (n = 10). In the latter, we conducted 20 closed-loop teleoperation trials where autonomous servoing was used. In 10 trials, the RCE traversed an area where large acoustic features were present. In the other 10 trials, smaller acoustic features were present. The travel distance (trajectory length) was 15 cm in all trials. We developed and validated a linear traversal strategy for the RCE where the system entered a μUS servoing mode when acoustic coupling was lost. Once coupling was recovered, the RCE attempted to maintain the orientation in which the high-amplitude echo, i.e., high ESR, was achieved. This methodology proved to be robust, and the majority of markers were identified. We were able to measure widths of markers and spacing between them with about 1-mm accuracy.

The intended outcome of these tests was to show the core functionality (i.e., acquiring μUS echoes using a magnetically actuated RCE) and was necessary to determine whether this technology should be explored further. Before the in vivo trial, the capsule was coated with a conformal film of Parylene C using a vacuum deposition tool (SCS PDS 2010, Specialty Coating Systems, IN, USA). The surface was primed with A174 silane adhesion promoter before deposition. Parylene C is a USP Class VI polymer that is commonly used for its biocompatibility, lubricity, and moisture barrier properties (64). This coating provided an additional layer of protection between the capsule and the surrounding tissue. Adhesion testing of Parylene C to the resin and leak testing of the capsule were also conducted before the in vivo trial, as detailed in the Supplementary Materials.

The trial was performed at a dedicated large animal research facility in the Roslin Institute, University of Edinburgh. The study was conducted under Home Office (UK) License (Procedure Project License: PF5151DAF) in accordance with the Animal (Scientific Procedures) Act 1986. The choice of a porcine as a model is based on GI anatomy comparable with human (73, 74). This includes bowel dimensions, particularly lumen caliber and bowel wall structure. The trial was conducted with the porcine (British Landrace, female, 37 kg) placed under general, terminal anesthesia. The system was used in an operating theater that had access to a mobile fluoroscopy machine (“C-arm”, Ziehm Vision FD). The C-arm was only used for visualizing the RCE in situ and was not necessary for its operation. Before trial initiation, the bowel was cleaned gently with water to remove any residual fecal material, and the ESR_thresh was calibrated in situ. This was the main source of feedback during tests and provided a quantitative measure of transducer-tissue contact to inform RCE motion.

The RCE was inserted per rectum and was advanced about 20 cm into the colon by a combination of pushing the tether and magnetic actuation. Air and water were administered through the irrigation channel when deemed necessary either to distend the bowel to advance the RCE or for improving transducer coupling. During teleoperated echo detection experiments, the RCE was first placed in a decoupled configuration. The test was then started, and the user increased the contact force and adjusted tilt within a predefined range until multiple distinct echoes were found or until no echo was found after more than 1 min of tilting motions. The RCE was then decoupled before completing the next repetition. During autonomous echo detection tests, the user placed the RCE in the colon and decoupled the transducer in the same way as the teleoperated experiments; however, in this case, the autonomous routine was started by pressing a button. The system then performed a search routine in which the RCE was tilted within a 10° range (−15° to −25°), and the force was increased. When an echo was detected, the system autonomously attempted to maintain the current configuration (tilting back and forth to compensate for environmental disturbances). If lost again, the search routine resumed. Movie S3 summarizes the in vivo work. Once the experiments were completed, the animal was euthanized, and necropsy was performed. There was no evidence of colonic perforation or gross trauma.