Exploring the proprioceptive potential of joint receptors using a biomimetic robotic joint
Authors: Akihiro Miki, Shun Hasegawa, Sota Yuzaki, Yuta Sahara, Yoshimoto Ribayashi, Kento Kawaharazuka, Kei Okada
Organizations: Mechano-Informatics, Graduate School of Information Science and Technology, the University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, 1138656, Tokyo, Japan
In neuroscience, joint receptors have traditionally been viewed as limit detectors, providing positional information only at extreme joint angles, while muscle spindles are considered the primary sensors of joint angle position. However, joint receptors are widely distributed throughout the joint capsule, and their full role in proprioception remains unclear. In this study, we specifically focused on mimicking Type I joint receptors, which respond to slow and sustained movements, and quantified their proprioceptive potential using a biomimetic joint developed with robotics technology. Results showed that Type I-like joint receptors alone enabled proprioceptive sensing with an average error of less than 2 degrees in both bending and twisting motions. These findings suggest that joint receptors may play a greater role in proprioception than previously recognized and that the relative contributions of muscle spindles and joint receptors are differentially weighted within neural networks during development and evolution. Furthermore, this work may prompt new discussions on the differential proprioceptive deficits observed between the elbows and knees in patients with hereditary sensory and autonomic neuropathy type III. Together, these findings highlight the potential of biomimetics-based robotic approaches for advancing interdisciplinary research bridging neuroscience, medicine, and robotics.
Figures & tables
Figure 1: Overview of the conceptual framework and methodology of this study. This figure illustrates how joint receptors were anatomically and functionally modeled, implemented in a biomimetic joint with embedded strain gauge sensors, and evaluated using a neural network inspired by the dorsal column–medial lemniscal pathway. a, An anatomical diagram of the human body model and joint capsule is shown (adapted from [ 25 , 26 ] ). b, The dorsal column–medial lemniscus pathway (DCML) is illustrated (adapted from [ 27 ] ). c, Based on anatomical structures, the joint capsule and joint receptors are modeled and implemented on an actual device. d, The dorsal column–medial lemniscus pathway is modeled using a neural network, which is trained to output proprioceptive joint position sense from joint receptor data obtained from the actual device. e, The verification is conducted to determine whether proprioceptive joint position sense has been acquired through training.
Figure 2: Design and implementation of the biomimetic joint used in this study. This figure illustrates how the biomimetic joint was designed with reference to anatomical structures, fabricated with embedded strain gauge sensors, and tested under different motion conditions (bending, twisting, and push–pull). a, Anatomical diagram of a biological joint. The joint capsule surrounds the joint formed by bone and cartilage. The anatomical figure is adapted from [ 29 ] . b, Design drawing of a biomimetic joint. The translucent yellow part represents the joint capsule. c, Various components of the biomimetic joint and the completed state after rubber molding. The 3D printed bone tips and cartilage parts are attached, and the soft joint capsule’s fibrous membrane is created using natural rubber latex to enclose them. d, Diagram showing the process of shaping the joint capsule while embedding strain gauge sensors, which act as joint receptors. The sensors are arranged in three rows, with the higher-numbered sensors placed closer to the bone attachment points. e, Image showing the biomimetic joint bent at approximately 90 degrees. f, Image showing the biomimetic joint twisted at approximately 45 degrees. g, Image showing the biomimetic joint being pushed and pulled by approximately 1cm .
Figure 3: Data acquisition and modeling framework for proprioception estimation using the biomimetic joint. a, The process of moving the biomimetic joint to acquire data from 60 strain gauge sensors embedded in the joint capsule, functioning as joint receptors, and joint position and orientation data. b, Joint movement data is obtained using SLAM Camera (Intel RealSense T265), providing relative coordinates and orientations from the initial posture. The mounting position of the camera and its coordinate system are shown. c, A deep learning model was developed to estimate the joint’s position and orientation from receptor data. The input consists of values from the 60 strain gauge sensors, and the output includes roll, pitch, yaw, and x, y, z translations. d-i, Visualization of changes in sensor values during movement (d–g), and corresponding orientation (h) and position (i). Sensor values are shown only for connected sensors and were read via a voltage divider circuit. The horizontal axis of all graphs represents time in seconds. The vertical axis of the sensor graphs corresponds to voltage readings from the circuit (with the top value of approximately 628 corresponding to 3.3 V), while those of the roll, pitch, and yaw graphs represent angles, and those of the x, y, and z graphs represent positions in meters. Annotations 1–6 correspond to panel (a).
Figure 4: Performance of proprioception estimation using the trained deep learning model. a-f, This graph shows the estimation of the joint’s coordinates and orientation generated using deep learning based on sensory receptor data collected from 60 strain gauge sensors embedded in the joint capsule. Due to the open-type ball joint loosely connected by soft tissues, the graph visualizes not only orientation(roll(a), pitch(b), and yaw(c)) but also translation(x(d), y(e), and z(f) positions), comparing them to actual measurements (ground truth). The vertical axis represents angles for roll, pitch, and yaw as well as position in meters, while the horizontal axis represents a dimensionless number denoting the number of test data points.
Error Type
x (m)
y (m)
z (m)
Maximum Error
0.0195
0.0149
0.0247
Mean Error
0.0023
0.0020
0.0033
Standard Deviation
0.0028
0.0022
0.0037
Table 1: Statistical Values on Positional Errors.
Error Type
roll (°)
pitch (°)
yaw (°)
Maximum Error
6.1621
7.8585
9.4371
Mean Error
0.7211
1.3025
1.3744
Standard Deviation
0.8798
1.4239
1.5882
Table 2: Statistical Values on Attitude Errors.
Figure 5: Redundancy analysis of proprioception estimation under reduced sensor input with ten independent trials. This figure shows the mean and standard deviation of angular errors across ten independent trials for different sensor reduction ratios. Panels (a) and (b) correspond to pitch and roll, respectively. Error bars represent the standard deviation across trials for each reduction ratio, magnified tenfold for visibility. Statistical analysis revealed that mean angular error first showed a statistically significant increase compared with the baseline at a reduction ratio of 25.7% for pitch and 28.6% for roll ( p<0.05 ). In contrast, maximum angular error exhibited statistically significant increases only at higher reduction ratios ( 48.6% for pitch and 51.4% for roll). These results indicate that while average performance degradation begins around a 30% reduction, extreme errors become prominent as the reduction approaches 50% . Notably, a mean angular error exceeding 2\SIUnitSymbolDegree was first observed at a reduction ratio of 45.7% in pitch.
Figure 6: Identification of key joint receptors for proprioception estimation. This figure highlights the sensory receptors that the deep learning model identified as most important during different joint movements, together with visualizations of their spatial distribution. Graphs showing key joint receptors important for the deep learning model during twisting (a), bending (b), and push-pull (c) movements near the limit of the joint’s range of motion, along with their respective visualization renderings (d corresponds to a, e to b, and f to c). Each receptor was labeled as “adc(i)_data(j),” where i corresponds to the board number and j to the sensor number, arranged in three rows from mid-capsular region as shown in Fig. 2 d. Color coding reflects both the frequency and magnitude of receptor importance across the five model-training trials. The receptor that appeared most frequently among the top three across trials is shown in red, the remaining one or two receptors within the top three are shaded according to their SHAP values, and all other receptors are displayed in dark gray.
Error Type
x (m)
y (m)
z (m)
Maximum Error
0.0209
0.0230
0.0276
Mean Error
0.0021
0.0023
0.0031
Standard Deviation
0.0033
0.0035
0.0047
Table 3: Table of Statistics on Positional Errors in Different Datasets.
Error Type
roll (°)
pitch (°)
yaw (°)
Maximum Error
7.9132
5.3688
8.6223
Mean Error
0.3172
0.6780
1.2479
Standard Deviation
0.7047
0.8202
1.6711
Table 4: Table of Statistics on Attitude Errors in Different Datasets.
Figure 7: Identification of disconnected strain gauge sensors. a, Plot of strain gauge sensors, including those with disconnections. For clarity, a selected subset of strain gauge sensors is plotted. b, Enlarged view of a portion of (a). The strain gauge sensor highlighted in orange within the dashed region is observed to be disconnected.
Figure 8: Validation of deep learning–based proprioception estimation with an additional dataset. This figure shows the performance of the trained model when tested on a dataset collected independently but with the same procedure as in the Performance Verification of Joint Receptors Using Deep Learning section. The biomimetic joint is an open-type ball joint loosely connected by soft tissues, which permits not only orientations but also translational movements. The predicted values are compared against actual measurements (ground truth). a-f, This graph illustrates the joint coordinates and orientation estimated using deep learning based on sensory receptor data obtained from 60 strain gauge sensors embedded in the joint capsule. The model was trained in the same manner as in the Performance Verification of Joint Receptors Using Deep Learning section but using a different dataset collected through the same method. Due to the open-type ball joint being loosely connected by soft tissues, the graph visualizes not only orientation(roll (a), pitch (b), and yaw (c)) but also translation(x (d), y (e), and z (f) positions), comparing them to actual measurements (ground truth). The vertical axis represents angles for roll, pitch, and yaw, as well as position in meters, while the horizontal axis represents a dimensionless number denoting the number of test data points. The results indicate a high prediction accuracy.
Figure 9: Training process for redundancy analysis. This graph illustrates the training process used to examine the redundancy of proprioception acquired through joint receptors, as discussed in the Analysis of Redundancy in Joint Receptors section. The horizontal axis represents the number of training iterations (epochs), while the vertical axis represents the loss function. The reduction ratio is represented by a continuous color gradient, as indicated by the color bar. As the reduction ratio increases, the training process exhibits slower convergence and higher final loss values.
Figure 10: Redundancy analysis of proprioception estimation under reduced sensor input from a single trial. This figure illustrates the angular errors for pitch (a) and roll (b) in one trial. Error bars represent the within-trial standard deviation of errors without scaling. Because this is a single trial, the results are subject to stochastic variation inherent in deep learning, and no statistical analysis can be performed. Nonetheless, visual inspection suggests a noticeable degradation in performance around a reduction ratio of approximately 50% , which is consistent with the averaged results presented in Fig. 5 .
Figure 11: Distribution of critical joint receptors in near-limit regions where twisting(roll) angles exceeded 20\SIUnitSymbolDegree . a-e, Bar graphs showing the results of five trials conducted in regions where twisting(roll) angles exceeded 20\SIUnitSymbolDegree in the Distribution of Critical Sensory Receptors Near Range of Motion Limits section. For each trial, SHAP values representing the contributions of all motion components (x, y, z, roll, pitch, and yaw) were aggregated to evaluate the overall importance of each sensory receptor. Among the top three most important receptors in each trial, 3, 6, and 6 were located in the mid-capsular, transitional, and bone attachment regions, respectively, indicating that receptors situated at or near the bone attachment were predominant. Notably, adc1_data7 and adc1_data12 were consistently included among the top three across all trials. These receptors are located in the bone attachment and transitional regions, respectively.
Figure 12: Distribution of critical joint receptors in near-limit regions where bending(pitch or yaw) angles exceeded 70\SIUnitSymbolDegree . a–e, Bar graphs showing the results of five trials conducted in regions where bending(pitch or yaw) angles exceeded 70\SIUnitSymbolDegree in the Distribution of Critical Sensory Receptors Near Range of Motion Limits section. For each trial, SHAP values representing the contributions of all motion components (x, y, z, roll, pitch, and yaw) were aggregated to evaluate the overall importance of each sensory receptor. Among the top three most important receptors in each trial, 0, 9, and 6 were located in the mid-capsular, transitional, and bone attachment regions, respectively, indicating that receptors situated at or near the bone attachment were predominant. Notably, adc0_data15 and adc1_data11 were consistently included among the top three across all trials. These receptors are located in the bone attachment and transitional regions, respectively.
Figure 13: Distribution of critical joint receptors in near-limit regions where the absolute value of x-translation exceeded 2mm . a–e, Bar graphs showing the results of five trials conducted in regions where the absolute value of x-translation exceeded 2mm in the Distribution of Critical Sensory Receptors Near Range of Motion Limits section. For each trial, SHAP values representing the contributions of all motion components (x, y, z, roll, pitch, and yaw) were aggregated to evaluate the overall importance of each sensory receptor. Among the top three most important receptors in each trial, 5, 6, and 4 were located in the mid-capsular, transitional, and bone attachment regions, respectively, indicating that mid-capsular receptors were also frequently identified as critical. Notably, adc3_data11 (transitional region) and adc2_data0 (mid-capsular region) were frequently included among the top three across the trials.
Proprioception in musculoskeletal humanoids is typically estimated primarily from muscle sensing, while the role of cutaneous deformation around joints remains insufficiently explored. In biological systems, mechanoreceptors distributed within soft tissue complement muscle feedback and support reliable joint state estimation. This study presents the design of a biomimetic joint-covering skin with a tissue-like layered structure that integrates pressure- and stretch-sensitive elements within the joint-covering tissue. The proposed skin is implemented on the musculoskeletal humanoid Musashi-W, and its independent proprioceptive capability as well as its integration with muscle sensing are evaluated. Experimental results show that the proposed skin alone achieves joint angle estimation with an average error of approximately 3 degrees. Furthermore, integration with muscle sensing improves estimation accuracy. Owing to its joint-covering structure, the skin may mechanically mitigate the influence of external disturbances on the muscles, and the integration of multiple modalities suggests the possibility of contributing to the identification of external stimuli that are difficult to interpret using muscle sensing alone. This work presents a design methodology for biomimetic joint-covering skin and demonstrates that such tissue-structured skin can serve as an effective approach for extending proprioceptive systems in musculoskeletal humanoids.
Department of Mechano-Informatics, Graduate School of Information Science and Technology, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-8656, Japan · AI Center, Graduate School of Information Science and Technology, The University of Tokyo, Japan
Maintaining stable internal representations of continuous variables is fundamental for effective robotic control. Continuous attractor networks provide a biologically inspired mechanism for encoding such variables, yet neuromorphic realizations have rarely addressed proprioceptive estimation under resource constraints. This work introduces a spiking ring-attractor network representing a robot joint angle through self-sustaining population activity. Local excitation and broad inhibition support a stable activity bump, while velocity-modulated asymmetries drive its translation and boundary conditions confine motion within mechanical limits. The network reproduces smooth trajectory tracking and remains stable near joint limits, showing reduced drift and improved accuracy compared to unbounded models. Such compact hardware-compatible implementation preserves multi-second stability demonstrating a near-linear relationship between bump velocity and synaptic modulation.
Federica Ferrari, Flavia Davidhi, Bernard Maacaron +6
Istituto Italiano di Tecnologia, Genoa, Italy · Department of Neurosurgery, University Hospital Zürich, Switzerland · Institute of Neuroinformatics, Univ. of Zurich and ETH Zurich, Switzerland +3
Micro-scale continuum robots face significant limitations in achieving three-dimensional contact force perception, primarily due to structural miniaturization, nonlinear mechanical, and sensor integration. To overcome these limitations, this paper introduces a novel proprioception method for cable-driven continuum robots based on proximal-integrated force sensing (i.e., cable tension and six-axis force/torque (F/T) sensor), inspired by the tendon-joint collaborative sensing mechanism of the finger. By integrating biomechanically inspired design principles with nonlinear modeling, the proposed method addresses the challenge of force perception (including the three-dimensional contact force and the location of the contact point) and shape estimation in micro-scale continuum robots. First, a quasi-bionic mapping between human tissues/organs and robot components is established, enabling the transfer of the integrated sensing strategy of tendons, joints, and neural feedback to the robotic system. Second, a multimodal perception strategy is developed based on the structural constraints inherent to continuum robots. The complex relationships among mechanical and material nonlinearities, robot motion states, and contact forces are formulated as an optimization problem to reduce the perception complexity. Finally, experimental validation demonstrates the effectiveness of the proposed method. This work lays the foundation for developing safer and smarter continuum robots, enabling broader clinical adoption in complex environments.
Gang Zhang, Junyan Yan, Jibiao Chen +1
Department of Mechanical and Automation Engineering and T Stone Robotics Institute, The Chinese University of Hong Kong, Hong Kong · Curve Robotics Limited, Hong Kong · Shun Hing Institute of Advanced Engineering and Multi-Scale Medical Robotics Center, Hong Kong