Introduction

Cochlear implantation (CI) is considered to be the treatment mainstay for hearing rehabilitation in patients with severe-to-profound hearing loss and limited benefit with amplification1,2. In clinical practice, after a retroauricular incision of skin, the mastoidectomy and posterior tympanotomy are made by a surgical drill to gain access to the cochlea. Then an electrode array (EA) can be inserted via the round window or cochleostomy. The surgeon generally uses special forceps to hold a tens to 34 mm3 long EA before inserting it into the scala tympani of the cochlea through a small, constrained tunnel of diameter of 2–5 mm4. However, the patient’s residual hearing and remaining inner ear structures are vulnerable to damage during the EA insertion process. The basilar membrane and spiral plate in the inner ear are extremely delicate and can be injured when subjected to an external force >100 mN5,6.

With the development of robot-assisted surgery research7, robotic CI (with precise and safe manipulation) has been a primary focus8,9,10 to reduce trauma and improve surgical efficacy. Researchers from Intuitive Surgical Inc. in Sunnyvale, California, employed a da Vinci Si surgical robot to conduct the first master-slave-assisted cadaveric CI under augmented visual navigation11. The RobOtol® system developed by Collin Medical received approval for clinical applications focused on assistance for cochlear implants12. The HEARO® robotic system introduced microprecision imaging and specialized robot-assisted drilling technology to reduce invasiveness of the procedure13. IotaSoft® systems (with Food and Drug Administration (FDA) approval) were designed to free up a surgeons’ hands and reduce variability and forces by controlling the speed of EA insertion14. Nonetheless, in all of these well-developed systems, the interaction forces between the EA and cochlea were not detected, which limits further improvements in the CI safety and efficacy6,15,16. Some scholars have used either commercial sensors or force-sensing elements attached to the end of the forceps, robot, or cochlear phantom, to measure the interaction forces during CI5,17,18. However, these methods do not achieve sensing at the EA forceps tip, and lack muti-axial force sensing capability, high sensitivity and resolution.

To achieve interaction force or tactile sensing technique, various methods have been developed based on the principles of magnetics19, resistance20,21,22, capacitance23,24,25,26, piezoelectricity27,28,29, vision30,31,32, and optics33,34,35, etc. Among them, force sensors based on the fiber Bragg grating (FBG) sensors exhibit high adaptability to robotic assisted surgery because of their advantages of small size, high sensitivity, and good stability36. On the millimeter-scale instruments for ophthalmic microsurgery, the FBG force sensors with sub-millinewton resolution were integrated37,38, which can surpass the force-sensing sensitivity and accuracy of the human hand. In endoluminal intervention and diagnosis, the FBG force sensors at the distal end of the slender instruments can enhance the efficacy and safety39,40,41. Moreover, the multi-axial force/torque (F/T) sensors base on FBG sensors also have been developed for neurosurgical instruments42, robot joint43, and surgical forceps44,45. However, these sensor designs still have limitations that do not meet the requirements of CI. Firstly, the interference of the triaxial torque on the triaxial force cannot be decoupled compeletly, which cannot ensure the sensing accuracy. Secondly, the resolution balance between different axes, temperature compensation and CI workspace adaptability need to be further considered. Thirdly, the existing force-sensing surgical forceps do not fully decouple the interference of forceps actuation on force sensing, especially the cable-driven forceps. The cable friction and forceps reactive force may reduce the accuracy of external force sensing. Using pneumatic or magnetic actuators46,47 are more suitable for this decouping requirement. Additionally, EA integrated with FBG sensors were also developed to directly sense the resistance during CI48, but these designs require modifications to the existing EA, which is not practical in clinical practice.

In this work, we aim to develop a compact serial-parallel robotic platform with a 6-axis force-sensing pneumatic forceps for CI of the straight electrode array. It can provide 6-DOF micromanipulation capability for surgical forceps with independent 3-DOF translations and 3-DOF rotations. The proposed surgical forceps has the ability to actively grip and release the EA and can sense 6-axis real-time F/T at the forceps tip during EA insertion. This combination can provide more convenient, accurate, and safe robotic CI. For the design requirements of the robotic platform, we propose a combination of a robotic arm and a spherical parallel mechanism (SPM). For the surgical forceps, we propose a combination of a pneumatic gripper and a curved 6-axis F/T sensor based on FBG sensors, introducing a hybrid parallel and inclined FBG sensor configuration attached to a curved substrate for F/T sensing with high and balanced resolution. This 6-axis F/T sensor not only has the function of temperature compensation, but also its sensing readings are not affected by the actuation state switching of the pneumatic gripper. That is, the actuation and F/T sensing of the surgical forceps are decoupled. A comparison between this study and existing robots and automated tools10,12,17,49,50,51 to assist EA insertion is provided in the Table. S7, which further demonstrates the advanced capabilities of the proposed robotic forceps. Subsequently, we built a complete robot prototype and validated its feasibility through a robotic CI study on a phantom. Furthermore, we collaborated with surgeons to conduct robotic CI studies on human cadaver, including an isolated temporal bone and a cadaveric human head. The EA insertion forces and torques in cadaver studies were recorded in real time, and visual feedback was provided to the surgeon for timely consideration. These cadaver studies results proved that the proposed CI robot can assist surgeons in inserting EA more accurately and safely according to their own wishes. And the different computed tomography (CT) scans of the implantation results also proved the preliminary feasibility of the proposed CI robot.

Results

CI robot concept design

We propose a CI robot concept design including a compact serial-parallel robotic platform and a force-sensing pneumatic surgical forceps. The deployment relationship between the patient, CI robot, and microscope is shown in Fig. 1A. As an example, to perform CI on the patient’s left ear, the microscope is placed above the patient’s ear. The surgeon usually sits behind the patient’s head, adjusting the microscope’s field of view to align the EA insertion path behind the auricle, and is ready to perform CI. Without changing the conventional deployment, the robot is placed on the upper side of the patient’s head, allowing the surgeon to quickly switch between manual and robotic CI. Additionally, the robot’s location next to the operating table does not change no matter the patient needs CI in the left or right ear, which improves the efficiency of preoperative positioning. And the robot can adapt to the environment without modifying the operating room because of its small footprint.

Fig. 1: Overview of the CI robot and surgical scenario.
figure 1

A Deployment of the robot, patient, and microscope. The microscope is positioned above the head of a side-lying patient, and the robot is placed on the upper side of the patient’s head to accommodate CI of the left or right ear. B Detailed composition: The designed surgical forceps comprises a curved F/T sensor and a pneumatic gripper. By setting the remote center to coincide with the pneumatic gripper center, the gripper’s independent 3-DOF translations and 3-DOF rotations can be provided by the robotic arm and SPM respectively during CI. C Surgical scenario: Illustration of the relationship between the designed surgical forceps and temporal bone anatomy when performing CI. The side view of temporal bone is shown above, and the sectional view is shown below. The EA is gripped by the pneumatic gripper, avoiding the facial nerve, and inserted into the cochlea through the round window.

Full size image

The robotic platform comprises a robotic arm and a SPM. The SPM moving platform has three rotational DOF around its spherical center point52, as all rotating pair axes in the mechanism intersect at that point, as shown in Fig. 1B. By adjusting this point away from the moving platform through a structural parameter design, the spherical center point can become a remote center, giving the moving platform the characteristics of remote center of motion (RCM). The vertical distance (\(d\)) between the remote center and the SPM moving platform is 141 mm. However, this design sacrifices the workspace of the SPM, so we designed its three rotating drive axes in a coaxial manner. Three motors are integrated to drive the three coaxial-driving arms through gear transmission. This design not only increases the rotation range of the SPM around the drive coaxial direction to ±180°, but also makes the SPM structure more compact.

The proposed surgical forceps, including a 6-axis F/T sensor and a pneumatic gripper, is connected to the SPM moving platform. By setting the pneumatic gripper center to coincide with the remote center, the robotic platform provides 6-DOF independent motion for the EA micromanipulation, including three rotations from the SPM and three translations from the robotic arm. This feature improves its accuracy and stability while adjusting only the orientation of the pneumatic gripper, and reduces the space occupied by the robot’s movement and the risk of collision between the robotic platform and its surroundings. The three redundant DOF of the robotic arm are not used during EA insertion but only for rapid rough positioning before surgery. As shown in Fig. 1C, the pneumatic gripper grips the soft straight EA, and maneuvers its body to pass through the facial recess and reach the round window, which is the entrance to the cochlea. It is critical to avoid collisions between the surgical forceps and the facial nerve to reduce secondary harm to the patient. Furthermore, in terms of reusability and sterilization, the robotic platform could be separated from the patient by disposable medical sterile protective drapes or covers. For the developed surgical forceps, low temperature hydrogen peroxide plasma sterilization is a feasible way to maintain reusable. Another method is that the gripper can be disposable, thereby extending the boundary of medical sterile protective drapes to the assembly interface between the gripper and sensor, which ensures reusability of the force sensor.

Force-sensing pneumatic forceps design

A surgical forceps that could directly sense F/T at the EA gripper was developed. As shown in Fig. 2A, the surgical forceps includes a curved 6-axis F/T sensor and a pneumatic gripper. The F/T sensor comprises a strain-sensitive substrate and four optical fibers, each with three FBG sensors. The curved substrate is a hollow cylindrical tube, with four optical fibers adhered in the reserved grooves on its outer surface. Two optical fibers are adhered completely parallel to the substrate centerline, defined as Fiber 2 and Fiber 4. The other two are parallel to the substrate centerline before the substrate starts to bend and become inclined after the substrate bends, defined as Fiber 1 and Fiber 3. Since a FBG sensor requires compression and tension to measure deformation, an appropriate distance needs to be reserved between adjacent FBG sensors to avoid mutual interference between their readings. Consequently, section \({s}_{1}\) is set at a distance \(l\) away from the substrate tip, and section \({s}_{2}\) is parallel to section \({s}_{1}\) at a distance a. Four FBG sensors on the four fibers are equally configured on the substrate with an angular interval of 90° on sections \({s}_{1}\) and \({s}_{2}\). Between sections \({s}_{2}\) and \({s}_{3}\), the substrate is bent at an angle of \(90^\circ -\beta\). Section \({s}_{3}\) is set at the position where Fiber 1 and Fiber 3 are inclined, and the inclined angle is \(\alpha\). The angular interval between these two inclined FBG sensors is 180°. Section \({s}_{4}\) is parallel to section \({s}_{3}\), and two paralleled FBG sensors on Fiber 2 and Fiber 4 are also configured symmetrically. The remaining detailed geometric parameters are marked in Fig. 2A and their specific values are provided in Table S2.

Fig. 2: Design and modeling of the force-sensing pneumatic surgical forceps.
figure 2

A Structure of the force-sensing pneumatic surgical forceps: a pneumatic gripper (green), and a sensor with four optical fibers and a curved substrate (orange). It shows the detailed structure: the whole assembly, and the configuration of optical fibers on different cross sections of the substrate. B Design of the pneumatic gripper. It includes a rigid shell and soft films to grip the EA. C Configuration of the inclined FBG sensors on the substrate and its mathematic model. D Relationship between the force/torque resolution, \(\alpha\) and \(\beta\). The left figure illustrates that \(\beta\) approximating 30° ensures force resolutions in all directions are close to one another. The right figure illustrates the torque resolutions while \(\alpha\) and \(\beta\) are changing. When \(\alpha\) is 45° and \(\beta\) is 30°, the torque \({m}_{z}\) resolution is close to the resolutions of \({m}_{x}\) and \({m}_{y}\). E Substrate assembly grooves and assembly errors. Upper: Continuous empty grooves on the substrate surface to assist the sticking of each optical fiber. Lower left: For the FBG sensor pairs not parallel to the substrate central axis. This pair of FBG sensors has a relative offset \(\varDelta \theta\) of the sticking angle. Lower right: For the FBG sensor pairs parallel to the substrate central axis. This pair of FBG sensors has a relative offset \(\varDelta \delta\) due to the axial assembly error.

Full size image

The pneumatic gripper is integrated at the F/T sensor tip, its working principle being as shown in Fig. 2B. The actuating air supply passes through the inner tube of the substrate to the chambers on both sides of the gripper. The soft films adhered to the chambers on both sides inflate and clamp the EA placed in the middle. When the actuating air pressure is reduced, the soft films deflate, and the EA can be removed from the top of the gripper. Considering the available literature concerning FBG sensor-based force sensing (which suggests a relatively low axial force-sensing resolution and sensitivity), we sought to improve the axial force-sensing capability of a muti-axial force sensor. Consequently, we adopted a curved substrate scheme to transfer the axial force in the transverse direction as a conventional straight substrate cannot respond to the axial strain effectively. The curved substrate also prevents the EA from being obscured by the sensor under the microscope view. Other innovations include the hybrid parallel and inclined configurations of the FBG sensor pairs, which can compute 6-axis F/T more effectively (compared to the parallel FBG sensors in triaxial force-sensing applications). Additionally, the symmetrical configuration can realize temperature compensation to reduce temperature disturbances, which can be a considerable factor in FBG sensor-based force sensing performance. The sensing requirements, including the range and resolution, can be satisfied by adjusting the structural parameter design, configuration of the FBG sensors, and stiffness of the substrate material.

Force/Torque estimation of fiber-optic sensor

When force and torque are exerted on the substrate, its deformation changes. However, it is difficult to measure small geometrical changes in the substrate directly. Consequently, FBG sensors are used to detect the substrate deformation changes. Moreover, computing of F/T can be realized by studying the relationship between the wavelength changes of FBG sensors and the deformation changes of substrate. The mapping relationship between the F/T and wavelength changes of the FBG sensors can be expressed as follows:

$${{{\boldsymbol{F}}}}=K\cdot \Delta {{{\boldsymbol{\lambda }}}}$$
(1)

where \({{{\boldsymbol{F}}}}={[{f}_{x} \quad {f}_{y} \quad {f}_{z} \quad {m}_{x} \quad {m}_{y} \quad {m}_{z}]}^{T}\) denotes the external 6-axis forces and torques, \(\Delta {{{\boldsymbol{\lambda }}}}={[\Delta {\lambda }_{1}^{{s}_{1}}\Delta {\lambda }_{2}^{{s}_{1}}\Delta {\lambda }_{3}^{{s}_{1}}\Delta {\lambda }_{4}^{{s}_{1}}\cdots \Delta {\lambda }_{1}^{{s}_{3}}\Delta {\lambda }_{2}^{{s}_{4}}\Delta {\lambda }_{3}^{{s}_{3}}\Delta {\lambda }_{4}^{{s}_{4}}]}^{T}\) denotes the wavelength changes in the 12 FBG sensors, and K denotes a 6 × 12 matrix that maps the relationship between \({{{\boldsymbol{F}}}}\) and \(\Delta {{{\boldsymbol{\lambda }}}}\). Based on the static equilibrium equations, we can derive the relationship between \({{{\boldsymbol{F}}}}\) in the sensor coordinate frame \(\{S\}\) and the curvatures (\({{{\boldsymbol{\kappa }}}}\)) of the substrate by analyzing four cross sections—namely, \({s}_{1}\), \({s}_{2}\), \({s}_{3}\), and \({s}_{4}\). The detailed derivation process is presented in Supplementary Text S1. The relationship between \({{{\boldsymbol{F}}}}\) and the curvatures (\({{{\boldsymbol{\kappa }}}}\)) can be mapped by matrix \({K}_{1}\) as

$${{{\boldsymbol{\kappa }}}}={K}_{1}\cdot {{{\boldsymbol{F}}}}$$
(2)

where \({{{\boldsymbol{\kappa }}}}={[{\kappa }_{{{{{\rm{s}}}}}_{1}x} \quad {\kappa }_{{{{{\rm{s}}}}}_{1}y} \quad {\kappa }_{{{{{\rm{s}}}}}_{2}x} \quad {\kappa }_{{{{{\rm{s}}}}}_{2}y} \quad {\kappa }_{{{{{\rm{s}}}}}_{4}x} \quad {\kappa }_{{{{{\rm{s}}}}}_{3}y} \quad {\kappa }_{{{{{\rm{s}}}}}_{3}z}]}^{T}\), \({K}_{1}\) is determined by the material and structural parameters of the developed sensor. The next step is to determine the relationship between the FBG wavelength changes (\(\Delta {{{\boldsymbol{\lambda }}}}\)) and the curvatures (\({{{\boldsymbol{\kappa }}}}\)). The wavelength changes of the FBG sensors can be calculated from the strains and temperature as follows:

$$\Delta {\lambda }_{i}^{{s}_{j}}/{\lambda }_{i}^{{s}_{j}}={k}_{\varepsilon }\Delta {\varepsilon }_{i}^{{s}_{j}}+{k}_{T}\Delta {T}_{i}$$
(3)

where \(\Delta {\lambda }_{i}^{{s}_{j}}\) denotes the wavelength change of the ith FBG sensor in the section \({s}_{j}\), \({\lambda }_{i}^{{s}_{j}}\) denotes the reflected wavelength of the ith FBG sensor in the section \({s}_{j}\), \(\Delta {\varepsilon }_{i}^{{s}_{j}}\) denotes the strain change from the deformation of the ith FBG sensor in the section \({s}_{j}\), \(\Delta {T}_{i}\) is the temperature change at the sensor location, \({k}_{\varepsilon }\) and \({k}_{T}\) denote the coefficients of strain and temperature sensitivity53. All the FBG sensors are adhered to the substrate surface and can sense only the axial strain along themselves. Consequently, it is necessary to convert the three-dimensional strains on the substrate surface at different cross sections to the corresponding axial direction of each FBG sensor. As shown in Fig. 2C, for the inclined FBG sensors in section \({s}_{3}\) of Fiber 1 and Fiber 3, the inclined angle (\(\alpha\)) can affect two curvatures (\({\kappa }_{{{{{\rm{s}}}}}_{3}y}\) and \({\kappa }_{{{{{\rm{s}}}}}_{3}z}\)) sensing performance. An analysis of the parallel and inclined configuration FBG sensors is explained in Supplementary Text S2 and S3 in detail. We can then obtain the relationship between the \(\Delta {{{\boldsymbol{\lambda }}}}\) and curvatures (\({{{\mathbf{\kappa }}}}\)) as follows:

$${{{\boldsymbol{\kappa }}}}={K}_{2}\cdot \Delta {{{\boldsymbol{\lambda }}}}$$
(4)

where \({K}_{2}\) is determined by the FBG sensor and substrate structure parameters. Owing to the symmetry of the structure and configuration, the relationship between the curvatures and wavelength changes of the FBG sensors is linear when subtracting each of the two symmetrically configured FBG sensors, which also realizes temperature compensation. Finally, the constant-matrix (\(K\)) in Eq. (1) can be derived by combining Eqs. (2) and (4), that is \(K={K}_{1}^{-1}{K}_{2}\),which can be calibrated experimentally.

Forceps-sensor performance analysis

After obtaining the mapping matrix (\(K\)), we can analyze the different substrate structural parameters that affect the F/T sensor resolution of each axis, including the curved angle (\(\beta\)) of the substrate and the configuration parameters of the FBG sensors. The left of Fig. 2D shows the relationship between the force resolution and \(\beta\), illustrating that \(\beta\) approximating 30° ensures force resolutions in all directions are close to one another. The right of Fig. 2D illustrates the resolutions of torques with changing \(\alpha\) and \(\beta\). When \(\beta=30^\circ\), the torque (\({m}_{z}\)) resolution is close to the resolutions of \({m}_{x}\) and \({m}_{y}\), and we then adopted \(\alpha=45^\circ\). A detailed analysis and the determination of other parameters to ensure the developed sensor exhibited a high, balanced resolution are summarized in Supplementary Text S4 and Table S1. The design results of the forces resolutions are 0.14 mN, 0.14 mN, and 0.15 mN, and the torques resolutions are 2.2 mN∙mm, 2.2 mN∙mm, 1.8 mN∙mm, respectively (When the wavelength change resolution is 1 pm).

Besides, since the FBG sensors are manually adhered to the substrate, assembly errors are unavoidable and may affect the performance of the sensor. To facilitate manual assembly, continuous grooves are designed into the substrate surface, as shown in Fig. 2E. Additionally, the spacing between the three FBG sensors on each optical fiber is customized to guarantee consistency between the actual assembly and the design. However, there possibly exists a manual assembly error that is most likely to affect sensor performance: positional deviation between pairs of sensors, which could result in incorrect curvature measurements of the same cross-section. This assembly error is analyzed in two cases, as shown in Fig. 2E. The detailed analysis process is provided in Supplementary Text S5. The results show that the impacts of this assembly error on sensing accuracy are slight. The actual calibration and validation results of the sensor are acceptable with the impact of manual assembly errors. Moreover, sensor performance may be affected by its substrate deformation under large loads, which is validated by the simulation in Supplementary Text S6. The results show that substrate deformation does not limit the application of sensor within our required scenario.

Sensor calibration and validation

The model in Eq. (1) reveals the relationship between the F/T and wavelength changes of the FBG sensors. However, the mapping matrix (\(K\)) is difficult to obtain directly and precisely, making it is necessary to conduct calibration procedures to calculate \(K\) accurately. The calibration procedures can be conducted using a semi-automatic calibration method, as shown in Fig. 3A. Forces are induced by loading weights onto the sensor while torques are generated by exerting weights using an eccentric distance on a crossroad beam as the force arm. Weights ranging from 0 mN to 100 mN are used as the F/T loads. When the sensor pose changes owing to the robot motion, the sensor F/T state can be changed. By recording every sensor pose, the corresponding weight, and the wavelength changes, a data set of the sensor F/T state and wavelength changes can be generated. Since temperature compensation is realized by the symmetry of every FBG pairs, and other environmental disturbances are assumed to be negligible, \(K\) can be obtained by linear regression of this dataset.

Fig. 3: Calibration and validation experiment setup and results.
figure 3

A Experimental setup and data flow of the calibration and validation experiments. B i–iii: Force validation results. Forces predicted using the calibrated model and the actual forces, and their estimated errors. iv–vi: Torque validation results. Torques predicted using the calibrated model and the actual torques, and their estimated errors.

Full size image

Results of the calibration procedures are listed in Table S3, and Fig. S4 shows the forces and torques predicted by the calibrated model and the actual forces and torques. The matrix \(K\) is recorded in Supplementary Text S7. The resolutions of the three axis forces \({f}_{x}\), \({f}_{y}\), and \({f}_{z}\) are 0.07, 0.09, 0.17 mN, respectively, and the resolutions of torques \({m}_{x}\), \({m}_{y}\), and \({m}_{z}\) are 0.99, 1.16, 1.55 mN·mm, respectively. The mean absolute errors of the forces are 0.80, 1.24, and 2.11 mN, respectively, and of the torques are 11.04, 23.90, 43.50 mN·mm respectively. Consequently, in a 100-mN force range of force and 500-mN·mm torque range, the sensor’s inaccuracy amounts to 1.61%, 2.48%, 4.21%, 2.21%, 4.78%, and 8.70%, respectively.

Validation experiments were conducted to evaluate the accuracy of the calibrated F/T sensor under a random external F/T. Using the same experimental setup as that used in the calibration experiments, the robot moved the sensor through random poses and the weights and corresponding wavelength changes were collected. Figure 3B shows the estimated F/T values versus the applied values and their estimated errors. The estimated forces and torques can be calculated using the calibration matrix \(K\) and the detected wavelength changes collected by the FBG interrogator. In total, 250 groups of sampling data were collected, the absolute mean errors of \({f}_{x}\), \({f}_{y}\), and \({f}_{z}\) being 0.95, 1.26, 2.32 mN, respectively, and the absolute mean errors of \({m}_{x}\), \({m}_{y}\), and \({m}_{z}\) being 14.12, 23.22, and 45.32 mN·mm, respectively. These results are summarized in Table S4. As the validation experiments show, the sensor developed in this study exhibited the desired sensitivity and accuracy for 6-axis F/T sensing.

Robot construction and phantom study

A prototype system for the subsequent experiments was built, its main components being as shown in Fig. 4A. A multi-joystick controller serves as an effective and simple interface, providing an intuitive and straightforward method for mapping robot movements. This controller allows surgeons to independently control each DOF of robot movements, making it easy to achieve the desired operation. PC1 serves as the master node and receives motion instructions from the controller to control the robotic platform while displaying the microscope view and F/T sensing graphs on the screen to the operator in real time. As the slave node, PC2 collects and solves the F/T information through a FBG interrogator and transmits it to PC1. The SPM is manufactured using a 3D printing technique and is attached to the robotic arm to manipulate the surgical forceps that grips the medical training straight EA. The RCM characteristics of SPM is shown in Movie S1. The F/T sensor substrate is 3D printed, with grooves on its surface designed for embedding custom optical fibers. There are three red markers of FBG sensors on each fiber, which must be aligned as correctly as possible when adhered to the substrate. The body of the pneumatic gripper is also 3D printed, the chambers on both sides being covered with 0.3-mm thick silicone films and sealed using an adhesive. The demonstration of gripping and releasing EA using this pneumatic gripper is shown in Movie S2. The coordinate frame \(\{R\}\) of the end effector is set at the center of the pneumatic gripper, coinciding with the remote center of the SPM, which differs from the F/T sensor coordinate frame. We furtherly use a feather to demonstrate the F/T sensing function more intuitively in Movie S3. Moreover, a preliminary evaluation of the SPM motion accuracy was also performed, and the details are available in the Supplementary Text S8 and Fig. S6. The results show that the maintaining accuracy of the SPM remote center is ±0.3 mm, and the SPM rotary repeatability is ±0.14°.

Fig. 4: Prototype and preliminary tests.
figure 4

A Left: Demonstration of the prototype main components. Middle: The spherical parallel mechanism (SPM) is attached to the robotic arm and maneuvers the surgical forceps that grip the EA. Right: Close up of the surgical forceps. The coordinate frame of the robotic platform end effector set at the center of the pneumatic gripper, coinciding with the remote center of SPM. B Illustration of the robotic EA insertion process from the perspective of the translucent cochlea phantom. The red dotted outlines in each figure indicate the position of the EA tip. C The impact of the gripper actuation on F/T sensing is negligible through the experiment. Left: The force sensing and torque sensing graphs during the experiment. Right: The gripper cycles between inflated and deflated states.

Full size image

After building the prototype, we performed CI on a temporal bone phantom. The EA used was the same as above, was not pre-curved and had no stylet or sheath. By observing the array of electrode contacts on the EA spiraling forward within the cochlear phantom (Fig. 4B and Movie S5), we could verify the adaptability and feasibility of the developed CI robot. We then conducted micro-X-ray scans of the phantom after EA insertion to clearly evaluate the insertion status, as shown in Fig. S7. Regrettably, the scan results showed that the EA was not fully inserted. There were two possible reasons for this. Firstly, the operator was not a surgeon and lacked CI experience. Secondly, the friction inside the cochlear phantom can be large, making it easy to buckle and difficult to move forward passively. However, this phantom study showed that people without surgical training could perform CI with robotic assistance, demonstrating the potential to provide rapid CI training.

Decoupling validation of robotic forceps

The combination of the pneumatic gripper and F/T sensor was intentionally designed to decouple actuation and sensing of the surgical forceps. We selected a polyimide (PI) tube of inner diameter of 0.1 mm as the actuating channel and set it in the F/T sensor center hollow channel. Theoretically, it does not produce large forces on the channel because the air pressure required for gripper actuation is low. Therefore, the impact on the F/T sensor is minimal. We further confirmed this hypothesis experimentally.

The left-hand side of Fig. 4C shows the F/T curves for the frame \(\{R\}\) over time. After initializing the F/T sensor and leaving it for approximately 10 s (Phase I), the gripper cycles between the inflated and deflated states at approximately 5-s intervals (Phase II). The right-hand side of Fig. 4C shows that the gripper’s soft films reach the appropriate expansion state under an air pressure of 0.2 bar. This decoupling validation is shown in Movie S4. The F/T curves show that the sensor force and torque readings maintain reasonable fluctuation ranges of ±2 mN and ±5 mN·mm respectively in both phases, demonstrating the credibility of the surgical forceps actuation and sensing decoupling design concept.

CI robot validation on temporal bone

After validating the feasibility of robotic CI through the phantom study, we conducted further human cadaver studies, including robotic CI on an isolated temporal bone and a cadaveric human head. The deployment of the robot prototype, experimental subject, microscope, and a simple human-machine interface are shown on the left of Fig. 5A. For convenient robotic control and F/T feedback for the surgeon, we designed a simple human machine interface that displayed the robot’s motion, microscope view, and F/T sensing curves in real time on the screen. Firstly, we conducted a robotic CI experiment using the human cadaver temporal bone as the subject (middle of Fig. 5A), the EA being identical to that used in phantom study. Its effective length was 22 mm, and it was gripped in a straight orientation, 35 mm from its tip.

Fig. 5: Human cadaver studies.
figure 5

A Left: Scene of the surgeon performing robotic CI. Middle: An isolated temporal bone serves as the study subject of robotic CI. Right: A human cadaver head serves as the study subject of robotic CI. B Key frames recorded from the microscope perspective during the temporal bone study. C Force sensing and torque sensing graphs during the temporal bone study. D After completion of the robotic CI, microscopic observation, and cone beam CT (CBCT) scans are used to evaluate whether the EA is completely inserted.

Full size image

The entire EA insertion process can be divided into five steps—that is, rough positioning between the robot and temporal bone, gripping EA, adjusting the microscope, inserting the EA, and releasing it. Before the EA is gripped, it needs to be straightened as much as possible. This ensures that the insertion motion can be effectively transmitted to the EA tip and allows the surgeon to better control the EA movements. To accurately measure the insertion force, the surgeon is expected to avoid the collisions between the EA and the anatomical structures other than the cochlea. The key frames of the temporal bone study and the F/T sensing graphs are shown in Fig. 5B, C respectively. The microscope view is focused on the round window, which is the entrance for the EA insertion. The surgical forceps become increasingly visible during the process, which helps the surgeon judge how deep the robot is moving. This process can be divided into three phases based on the surgeon’s operation, taking approximately 168 s in total. During Phase I, the surgeon controls the robot, adjusting the EA tip to reach the round window in the correct orientation. The gravity and inertia forces of the EA cause the F/T readings to fluctuate over a small range. In the Phase II, the surgeon begins inserting the EA. During the first 10 s, the F/T readings remain unchanged owing to the thin, soft EA tip. The force \({f}_{z}\) then shows semiperiodic fluctuations since the surgeon controls the insertion in a step-by-step manner, and its amplitude gradually increasing to + 25 mN. This insertion manner ensures the robot movements can be effectively transmitted to the EA tip, avoiding the EA buckling outside the cochlea. Referring to our previous surgical experience, the surgeon usually pauses for a while after each step of inserting the EA to allow its own elasticity to relieve the advancing resistance. This enables the EA to “wriggle” toward the cochlea in a manner similar to an inchworm. With each step of EA insertion (the robot starts and stops once), the force readings increase and decrease to form a local peak. Therefore, this force readings also reflects the action of surgeon controlling the robot to insert EA. The surgeons can then transfer their clinical experience to perform robotic CI, demonstrating the effectiveness of the F/T sensor and CI robot. During the phase III (the insertion of the last six EA electrode contacts), additional insertion force (\({f}_{z}\)) is applied to overcome the resistance within the cochlea, which also produces a larger torque (\({m}_{y}\)). The maximum force (\({f}_{z}\)) and torque (\({m}_{y}\)) are approximately + 100 mN and + 240 mN·mm, respectively, which are close to the cochlear implant F/T safety limits.

After the surgeon determines that further EA insertion is neither safe nor effective, the EA is released. Here, the microscope view is magnified to obverse if there are any uninserted electrode contacts. To our satisfaction, there were no exposed electrode contacts, as shown in Fig. 5D (the left-hand side). Further cone beam computed tomography (CBCT) scans confirm this result (the imaging set up is shown in Fig. S8), the bright white spots connected in a spiral pattern, as shown in Fig. 5D (the right-hand side), being the electrode contacts on the EA. The processes and results of this temporal bone study are also shown in Movie S6.

CI robot validation on cadaver head

Finally, we conducted a robotic CI experiment using a cadaveric human head as the subject, as shown on the right side of Fig. 5A, The experimental procedure and layout being consistent with those used in the temporal bone study. The speed of the robot movements was lowered to 0.5 mm s-1 based on the surgeon’s advice. (It was 2 mm s-1 in the temporal bone study). This process could be divided into three phases, taking approximately 161 s in total (Fig. 6A, B). Compared to the F/T graphs of the temporal bone study, the time required for the EA tip to reach the round window in Phase I was shortened. This could be due to the improvements in the surgeon’s operating experience. The F/T fluctuation amplitude in Phase II was smaller and smoother owing to the reduced the speed of robot allowing the surgeon to control the robot for continuous EA insertion without waiting for the EA advancing resistance to be released. Another possible reason could be that the interior of the cochlea in cadaver head was more lubricated, reducing the EA advancing resistance. As shown in Fig. 6B, the maximum force (\({f}_{z}\)) and torque (\({m}_{z}\)) during Phase III, are approximately + 60 mN and + 150 mN·mm, respectively, which are in a safer range than those in the temporal bone study. We used the similar evaluation method used for the temporal bone study, as shown in Fig. 6C, the results demonstrating that the EA was fully inserted into the cochlea, proving the preliminary feasibility of the proposed CI robot for safely and accurately performing CI in clinical practice. The processes and results of this cadaver head study are also shown in Movie S7.

Fig. 6: Cadaver head study.
figure 6

A Key frames recorded from the microscope perspective during the human cadaver head study. B Force sensing and torque sensing graphs during the human cadaver head study. C After completion of the robotic CI, microscopic observation, and high-resolution CT (HRCT) scans are used to evaluate whether the EA is completely inserted.

Full size image

Discussion

In this study, we developed a robotic forceps with decoupled pneumatic actuation and 6-axis force sensing for cochlear implantation. The robotic forceps can overcome the difficulty of manual precision surgical operations and the lack of multi-axis force-sensing capability during EA insertion. The 6-DOF robotic arm and the customized SPM together construct the entire motion platform. The SPM has mechanical RCM characteristic to perform 3-DOF rotations for the forceps. In clinical practice, the surgeon usually needs to manipulate proximal end of the forceps to achieve tilting or rotating around its distal end. Therefore, the distal rotation center on the forceps can be considered as the remote center. The SPM can effectively meet the requirements of these surgical operations. We use the three redundant DOFs from the robotic arm for robot preoperative rough positioning, and disabled them during EA insertion. So, the whole robotic platform does not have redundancy during EA insertion. Compared with the solution only using a single robotic arm, our developed robotic platform provides independent 3-DOF translations and 3-DOF rotations for the forceps. This independent configuration makes the translational and rotational workspaces independent of each other, and ensuring precision and stability in the robotic EA insertion.

The CI surgical forceps included a curved 6-axis F/T sensor and a pneumatic gripper. We designed the F/T sensor based on a curved substrate, using a hybrid parallel and inclined configuration of the FBG sensor pairs. Specifically, we used four optical fibers, each inscribed with three FBG sensors, to sense the strain changes on the surface of 3D printed substrate. We then built a static model to demonstrate the linear relationship between the external F/T and wavelength changes in the FBG sensors. For the EA gripper, it can actively grip the EA by inflating the soft films on both sides using air pressure. And we attached this gripper to the end of the F/T sensor. Finally, we verified the decoupled relationship between the gripper actuation and F/T sensing experimentally. From the experimental results in our studies and CI clinical requirements, the surgeons pay more attention to the insertion force feedback, rather than the torque. But if we only use a triaxial force sensor, the unmodeled torque also causes the wavelength changes of FBG sensors, leading to an incorrect force estimation. So, the torque measurements are still necessary, and we carried out the further analysis and simulation verification in the Supplementary Text S9. Besides, we also discuss the potential impacts of the gripper’s soft film buffering effect and gripper itself on the effectiveness of EA insertion force measurement in the Supplementary Text S10. The results show that their impacts are slight and negligible. Therefore, the force measurements of the developed sensor can effectively reflect the EA insertion force.

In the human cadaver studies, the surgeon first conducted the cadaveric temporal bone study using the developed robotic forceps with a moving speed of 2 mm/s, and then adjusted the speed to 0.5 mm/s to complete the cadaveric head study. As shown in the videos of these two cadaveric studies (Movie S6 and S7), the EA does not make unexpected contact with the facial recess or other anatomical structures of the temporal bone. Therefore, the insertion force measurement results are reliable. The insertion forces of both studies gradually increase, but the former is accompanied by relatively high-frequency vibrations, while the latter is smoother. The reason is that the surgeon uses different insertion methods (step-by-step movements or continuous movements) according to the speed. If the surgeon uses a step-by-step insertion method at high speed, the robot will frequently accelerate and decelerate, causing vibration. The results show that using the insertion method of continuous movements combined with the robot’s low-speed motion can perform more ideal EA insertion (insertion force increases slowly and steadily within safety range). Moreover, we conducted more comparisons with different insertion speeds and robotic/manual insertion in the Supplementary Text S11.

In both cadaveric studies, the surgeon fully implanted the EA and praised the robot’s micromanipulation and millinewton-level force sensing capabilities, indicating that the developed robotic forceps can assist the surgeon to complete CI. Many previous works have also performed robotic insertion forces monitoring using commercial or custom force sensors10,54,55. Compared with their insertion force results, our results from both cadaveric studies show a good consistency with them in terms of maximum insertion force. Clinically, surgeons pay much attention to the hearing outcomes improvements56 of reducing EA insertion trauma by using robot-assisted method. To understand the intracochlear trauma formation from the EA insertion, it is significant to study the behavior and status of EA at different insertion depths. Intracochlear pressure and insertion force changes are two main indicators for analyzing the EA insertion process56,57. And the insertion speed is one of the key factors affecting these two indicators. A slower insertion speed can significantly reduce the peak value of intracochlear pressure during EA insertion58,59,60. In addition, in terms of the insertion force, a slower insertion speed generally reduces the force, but once the speed falls below a certain threshold, the insertion force does not decrease further54,55,61.

Moreover, the correlation between the peak insertion force and the level of potential intracochlear trauma56,57 is not fully investigated thoroughly, where the insertion force is not the only factor to affect the trauma. The clinical experience for predicting or controlling potential trauma also involves the EA movement monitoring inside the cochlea. However, real-time monitoring of EA movements is difficult, so it is potential to estimate EA actual movements via the real-time insertion force combined with the prior model of EA movements. Some previous studies have modeled the friction inside cochlear62,63, which may be effective to establish this prior model (a correlation between insertion force, insertion depth and EA movements). The robot we developed has a good motion accuracy and a satisfactory force sensing capability, that can help build this prior model in the future. Currently, our robot operates at a constant speed during EA insertion, and the surgeon only controls the start and stop of the robot movement. But in the future, based on the estimated results of EA movements, the robot could dynamically adjust its motion, including speed and force, to minimize the intracochlear trauma. Besides, electromagnetic navigation system64 or electrocochleography (ECochG) responses65 could be integrated into the robot control. Furtherly, a robot-assisted EA insertion standard could be established to ensure the consistency in limiting intracochlear trauma.

However, the developed robotic forceps still have some shortcomings. First, the EA gripper needs further miniaturization that can enter the facial recess. When the round window is in a biased position, making it difficult to insert the EA, the EA gripper should still allow the surgeon apply the appropriate insertion force more efficiently. Second, when the EA initial insertion is incomplete, it is difficult for the gripper to readjust the gripping position on EA, which limits the practicality of the robot. Third, the F/T sensing feedback needs more direct—that is, it can be amplified and fed back to the surgeon’s touch or displayed under the microscope view. Fourth, the F/T sensor design can have difficulties with clinical disinfection, and the exposed optical fibers on the substrate needs protection to ensure stable performance over multiple use. Fifth, our CI robot is not applicable for the EA insertion with a stylet/sheath, but we will improve it in future with a custom end-effector or with a dual-arm robot. Finally, from the perspective of clinical practice, surgeons usually use another instrument by their non-insertion hand to help guide the EA through the facial recess during insertion, or to stabilize the EA during releasing it. Therefore, we will extend the current version to a dual-arm robot, which can mimic the bimanual operations in the current clinical practice. And it also has better collaborative capabilities to achieve more complex operations, which is also the focus of future work.

In current microsurgery robotic research, the robot’s accuracy of movement and stability can easily surpass those of the human hand. However, challenges remain in sensing the F/T that interact with patients’ tissue or anatomical structures. The developed robotic forceps could provide the interactive F/T sensing for surgeons during CI micromanipulation and improve robotic surgery safety. It could also speed up surgeon training and make it more effective, allowing more patients to receive treatment aimed at improving their quality of life in a timely manner.

Methods

Ethical approval statements

Ethical approval was obtained from the Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine Ethics Committee (Shanghai, China). The study was registered with approval number SH9H-2023-T239-1.

Surgical forceps fabrication

The proposed F/T sensor comprised a substrate made of red wax (Young’s modulus 3.84 GPa, 3D printing, RC90, EnvisionTEC-P4K, USA), and four optical fibers with three FBG sensors inscribed on each (center wavelength 1540, 1550, and 1560 nm, diameter 125 µm, single 1-mm grating, Beijing Technica Inc., China). Optical fibers were manually adhered into the reserved substrate grooves using the medical adhesive (Loctite 4013, Henkel, Argentina). The pneumatic gripper comprised a rigid shell fabricated using a 3D printer (MicroArch S240, BMF Precision Tech Inc., China), and the soft films were made of silicone (Ecoflex 0030, Smooth On, Inc., USA). To obtain a uniform film of the target thickness, a special thickness preparer (Qigong Instruments, China) was used. The silicone sample plate was placed on a flat platform and fixed such that it did not move when the preparer slid. The preparation device was then set based on the required wet film thickness, placed on the short side of the sample plate close to the parallel or specified position, poured in front of the preparation device, and slid slowly to obtain a silicone film of the desired thickness. Two 0.4 × 0.6 mm films were then cut and adhered to the chambers of the 3D printing shell using plastic and rubber instant adhesive (PR40, 3 M, USA). A polyimide (PI) tube with an outer diameter of 0.1 mm is assembled as the pneumatic transmission tube for the gripper, and fixed by the medical adhesive (Loctite 4013, Henkel, Argentina).

Prototype construction

As shown in Fig. 4A, the robotic platform comprised a robotic arm (CR3 Dobot-robot Technology Co., Ltd. China) and a SPM driven by three piezo motors (LR23-50, PiezoMotor, Sweden). A joystick (XBOX One S, Microsoft, USA) was used as the operator controller. An overall control framework was developed on PC1 based on the robotic operating system (ROS). All the optical fibers were connected to an interrogator (FBG-Scan 908, FBGS International, Belgium), and the reflected light wavelength change in each fiber was interrogated by PC2 and sent to the PC1 via a local area network. The view of the microscope (MSD203, Murzider, China) also transferred to PC1. The human-machine interface developed was based on a simulated platform (Coppeliasim Edu, Coppelia Robotics, Ltd, Zürich, Switzerland). Matlab 2023 (MathWorks, Natick, USA) was used for analysis the data, and Solidworks 2021 (Dassault Systèmes, Versailles, France) was used for robot CAD design and illustration. The air pressure required to inflate the gripper is precisely measured by a microfluid controller (OB1 Mk4, Elveflow, France). And during the experiments, the air source of the pneumatic gripper is a small air pump, and the supply air pressure is adjusted by a pressure regulating valve. The EA we used in this study is a silica gel micro-EA (CS-10ATM, Nurotron, Hangzhou, China), which is a slim straight electrode array with 24 stimulus sites distributed on a 22 mm long silicone carrier. The 3D-printed temporal bone model used in this study is manufactured by Phacon GmbH (Leipzig, Germany)66. It was developed based on data acquired by micro-CT images of normal human cadaveric temporal bone (Patient “Schmidt”). The bone is constructed using a cast powder and a bonding agent that has similar material properties to bone. Colors are added to different anatomic structures to increase the realism of the model, and the cochlea is a transparent hollow structure. The surgeon performed a mastoidectomy and posterior tympanotomy on the model following a standard surgical procedure, and simulated the fluid environment by infusing physiological saline into the cochlea through the round window. The fixation tools of the temporal bone and cadaveric human head are a magnetic locking clamp and a custom head stabilizer respectively. The computed tomography (CT) equipment (LARGEV, Ultra3D) used to evaluate both cadaveric studies. A fresh frozen (Asian adult) cadaver head was used for cadaveric head study. Mastoidectomy and posterior tympanotomy were performed by a senior surgeon following a standard surgical procedure, and the round window was exposed before this cadaveric head study. All robotic CI procedures on the human cadaver temporal bone and head were performed by a chief physician with 15 years of experience in otological surgery.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.