An Enhanced Proprioceptive Method for Soft Robots Integrating Bend Sensors and IMUs

Dong Heon Han, Mayank Mehta, Runze Zuo, Zachary Wanger, Daniel Bruder · 2025

Read An Enhanced Proprioceptive Method for Soft Robots Integrating Bend Sensors and IMUs by Dong Heon Han, Mayank Mehta, Runze Zuo, Zachary Wanger, Daniel Bruder (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-01165.txt.

Introduction

Dong Heon Han and colleagues improve how soft robots sense their own shape, using only cheap, off the shelf sensors. Their method pairs inertial measurement units, which drift over time, with bend sensors that measure curvature directly, so each covers the other's weakness.

A Kalman filter fuses the two streams into segment tip orientations, correcting drift in a mutually compensatory way. A piecewise constant curvature model then reconstructs the robot's full deformation from those orientations. The robot ran for forty five minutes, under no load, external forces, and passive contact with obstacles. The shape error was about seventeen millimetres. That is under three percent of the robot's length, and a fifty six percent cut versus inertial sensors alone.

Catalog record

CREATORDong Heon Han, Mayank Mehta, Runze Zuo, Zachary Wanger, Daniel Bruder
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-01165

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

An Enhanced Proprioceptive Method for Soft Robots Integrating Bend Sensors and IMUs Dong Heon Han, Mayank Mehta, Runze Zuo, Zachary Wanger, Daniel Bruder Abstract This study presents an enhanced proprioceptive method for accurate shape estimation of soft robots using only off-the-shelf sensors, ensuring cost-effectiveness and easy applicability. By integrating inertial measurement units (IMUs) with complementary bend sensors, IMU drift is mitigated, enabling reliable long-term proprioception. A Kalman filter fuses segment tip orientations from both sensors in a mutually compensatory manner, improving shape estimation over single-sensor methods. A piecewise constant curvature model estimates the tip location from the fused orientation data and reconstructs the robot’s deformation. Experiments under no loading, external forces, and passive obstacle interactions during 45 minutes of continuous operation showed a root mean square error of 16.96 mm (2.91% of total length), a 56% reduction compared to IMU-only benchmarks. These results demonstrate that our approach not only enables long-duration proprioception in soft robots but also maintains high accuracy and robustness across these diverse conditions. I Introduction Soft robots possess intrinsic compliance and virtually infinite degrees of freedom, enabling continuous deformation [ 1 ] . To leverage these characteristics for safe and adaptive interaction in dynamic environments, proprioception—the ability to sense body configuration—is essential [ 2 ] . However, achieving accurate proprioception remains challenging due to material flexibility and complex nonlinear deformation [ 3 ] . Reliable proprioception is thus critical for enabling stable motion planning and control in soft robotic systems. Various transducers—resistive [ 4 ] , capacitive [ 5 ] , inductive [ 6 ] , magnetic [ 7 ] , and optical [ 8 ] —have been used for shape awareness in soft robots, often integrated into electronic skins [ 9 ] . Yet, these require specialized fabrication, limiting accessibility. Recent works show that off-the-shelf sensors can offer simpler, low-cost alternatives for shape estimation. Among such sensors, IMUs are popular for their compactness, ease of integration, and real-time tracking. Martin et al. [ 10 ] combined multiple IMUs with an extended PCC model [ 11 ] , identifying IMU drift as a limitation. Our method introduces a bend-sensor reference to correct IMU yaw drift before Kalman fusion, improving long-duration stability. Peng et al. [ 12 ] validated a tendon-driven manipulator with multiple IMUs but did not address drift. Stella et al. [ 13 ] used a kinematics-based filter with IMU orientation data but assumed negligible twist, restricting motion to planar bending. While this simplification reduces computation, it limits robustness under torsional impact, unlike our method which handles coupled bending–twisting deformation. Transient accelerations further amplify IMU noise and drift accumulation [ 14 ] , degrading accuracy in long-term operation. Fig. 1: Overview of our proposed proprioceptive sensing method for soft robots using off-the-shelf IMUs and bend sensors. Shape estimation is based on a piecewise constant curvature model using segment-tip orientations. IMU drift is corrected using bend-sensor references before fusion, enabling 45-minute continuous operation. To mitigate drift, previous studies fused IMUs with cameras [ 15 ] or draw-wire sensors [ 16 ] . However, camera-based fusion is affected by occlusion and lighting, and draw-wire methods require constant tension, introducing errors under compression. While such methods improve short-term accuracy, few demonstrate stable long-term drift correction under dynamic loads. Bend sensors offer an attractive alternative [ 17 ] , providing direct deformation measurement unaffected by lighting or tension. Although noisy and hysteretic, they exhibit minimal drift and insensitivity to acceleration spikes, capturing real structural changes. Yet, off-the-shelf bend sensors have not been systematically exploited to suppress IMU drift or sustain long-duration proprioception. Our approach addresses this gap by using bend sensors as low-frequency references that anchor IMU drift within a two-stage fusion pipeline for reproducible, long-term accuracy. The Kalman filter (KF) remains a standard tool for real-time sensor fusion, combining probabilistic modeling with prediction-correction cycles [ 18 ] . It is particularly effective for complementary fusion, integrating heterogeneous sensors [ 19 ] . However, its performance depends on accurate noise and covariance initialization, which is difficult in soft robots due to nonlinear material behavior [ 20 ] . Empirical calibration and data-driven approaches have therefore been used to model sensor deformation and improve estimation accuracy [ 21 , 22 ] . Building on these insights, we empirically tune KF noise and covariance parameters to capture soft-robot nonlinearities, enhancing proprioceptive performance. We propose a shape estimation framework using two off-the-shelf sensors: an IMU (BNO055, Bosch, Adafruit) and a bend sensor (2.2" Flex Sensor, Spectra Symbol), as shown in Fig. 1 . Drift correction is first applied by comparing IMU and bend readings, followed by KF fusion for refined estimation. A PCC model maps orientation to shape, and KF parameters are optimized by minimizing RMSE against ground truth. We validate the approach on a 2D planar manipulator, where yaw drift—parallel to gravity—is most pronounced, offering a controlled test before generalizing to 3D environments. This paper’s main contributions are threefold. First, a compensatory drift-correction scheme anchoring IMU yaw to a calibrated bend-sensor curvature reference. Second, a two-stage fusion of segment-tip orientation and PCC-mapped position using linear KFs, improving accuracy without added hardware. Third, a fully off-the-shelf, low-cost implementation validated for 45 minutes under no-load, load, and obstacle-contact conditions, achieving a 56% RMSE reduction versus an IMU-only PCC baseline while retaining real-time, embedded feasibility. The Methods section details the sensing and fusion design, the Experiments section presents validation procedures, and the Discussion and Conclusion interpret results and suggest future directions. Fig. 2: Experimental setup for validating the sensing method. (A) Planar robot test rig. (B) N th link with two actuators, four bend sensors (yellow), and two IMUs (blue) located between parallel sensors. (C) Full robot showing fiducial markers and sensing points for output comparison. II Methods A sensing segment is defined as the smallest unit in which multiple sensors collect data for deformation estimation. Each segment independently estimates its local state from these measurements. Proprioception is achieved using off-the-shelf IMUs and bend sensors to estimate the orientation of each segment endpoint, as illustrated in Fig. 2 . One end of the robot is fixed, and the structure bends laterally within a planar workspace. The bend sensor’s resistance varies with curvature, exhibiting a bijective relationship between endpoint orientation difference and measured resistance determined through calibration ( V–Orient Map in Fig. 4 ). To reduce noise and improve stability, two bend sensors are placed in parallel per segment, and their averaged output is used for orientation estimation, yielding both bend-sensor–mapped and IMU-derived orientations obtained from the IMU’s onboard filter. As shown in Fig. 4 , the sensing architecture integrates these measurements to estimate each segment’s orientation and position. IMU drift is first corrected using the bend-sensor reference ( IMU Correction ), and the corrected orientations are applied to the PCC model ( PCC ) for position estimation. Finally, the sensor outputs are fused through two Kalman filters— \textit{KF}_{\textit{orient}} for orientation and \textit{KF}_{\textit{coord}} for position—enhancing accuracy, suppressing noise, and ensuring stable proprioceptive performance. Fig. 3: A PCC kinematic model is used to map sensor readings to an estimate of the robot’s shape. The robot is assumed to consists of a chain of rigid and constant curvature segments. Fig. 4: Flowchart showing enhanced proprioception. The process starts

Read the complete text