Autonomous Robotic Palpation and Abnormality Detection through Ergodic Exploration

Autonomous Robotic Palpation and Abnormality Detection through Ergodic Exploration Luca Beber1,3,4, Edoardo Lamon3,4, Matteo Saveriano2,4, Daniele Fontanelli2,4, Luigi Palopoli3,4 Abstract—Palpation is a fundamental diagnostic tool for identifying stiffness variations in soft tissues, but its effectiveness is limited by subjectivity and practitioner variability. This work presents an autonomous palpation framework that continuously reconstructs viscoelastic tissue properties using only a standard force/torque sensor. The framework integrates forcebased viscoelastic estimation, Gaussian Process Regression to generate probabilistic stiffness maps, and ergodic trajectory planning to balance exploration and refinement adaptively. Both simulations and phantom experiments confirm the robustness of the approach, demonstrating accurate stiffness reconstruction and reliable inclusion localisation. These results highlight the potential of ergodic palpation as an objective and practical diagnostic tool in medical robotics. Index Terms—robotic palpation, ergodic exploration, tissue stiffness mapping, medical robotics I. INTRODUCTION Palpation enables clinicians to detect abnormalities, such as tumors or nodules, by assessing differences in tissue stiffness. While inexpensive and fast, manual palpation is inherently subjective, with diagnostic accuracy strongly depending on the practitioner. Imaging techniques, such as ultrasound and mammography, complement palpation but can be costly, timeconsuming, or unavailable in certain contexts. Robotic palpation offers the opportunity to make this process objective and repeatable. Previous work has relied on dense probing grids or Bayesian optimisation to guide probing [1], but these approaches are limited by efficiency, reliance on specialised sensors, or difficulties in delineating inclusion boundaries. In this work, we propose an ergodic palpation framework that continuously maps viscoelastic tissue properties while autonomously exploring the surface. Using only an off-theshelf force/torque sensor, the robot estimates elasticity and viscosity, updates a probabilistic stiffness map, and generates trajectories that balance global exploration with local refinement. An additional segmentation step provides accurate localisation of inclusions, yielding information that is directly relevant to clinical practice. We acknowledge the support of the EU project MAGICIAN (Grant Agreement n. 101120731). 1PhD program DRIM, University of Genoa. 2Department of Industrial Engineering, Universit` a di Trento, Trento, Italy. 3Department of Information Engineering and Computer Science, Universit` a di Trento, Trento, Italy. 4Interdepartmental Robotics Labs (IDRA), University of Trento. [email protected] II. METHODS A. Force-Based Viscoelastic Estimation To estimate the viscoelastic properties of soft tissue, a method called the Dimensionality Reduction Method (DRM) was employed [3]. In this approach, the 3D contact problem is projected onto a 2D space, resulting in a force that is proportional to the contact area. An Extended Kalman Filter (EKF) is then used to estimate penetration depth, local stiffness, and viscosity in real time, even during continuous motion [3]. B. Probabilistic Stiffness Mapping Each palpation provides local estimates of elasticity that are used to update a Gaussian Process Regression (GPR) model. The GPR produces a continuous stiffness map of the palpated surface, returning both predicted values and uncertainties. High variance highlights unexplored areas, while gradients of the mean stiffness reveal boundaries between soft and stiff regions. C. Ergodic Trajectory Planning Exploration is guided by the Heat Equation Driven Area Coverage (HEDAC) ergodic controller [2]. Unlike point-topoint probing, ergodic exploration generates smooth trajectories that allocate more time to regions of higher expected information. The target distribution is computed from a combination of stiffness uncertainty and boundary gradients, encouraging both exploration of unknown areas and refinement of inclusions. The process terminates when an ergodic convergence metric falls below a threshold, ensuring adequate coverage without unnecessary probing. D. Segmentation of Stiff Inclusions Once a stiffness map is reconstructed, inclusions are segmented by clustering regions of higher elasticity. A boundary extraction step converts these clusters into polygonal contours. The method is evaluated with standard metrics such as sensitivity (ability to detect inclusions), specificity (avoidance of false positives), and Intersection over Union (IoU) between estimated and ground-truth boundaries. High sensitivity ensures that no pathological region is missed, while good IoU indicates accurate delineation. III. EXPERIMENTS A. Simulation Studies Synthetic stiffness distributions with one to three inclusions were generated to test robustness. Across multiple runs with random initial positions, the algorithm consistently detected 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629722 97 t=14 (s) t=28 (s) t=42 (s) t=56 (s) t=70 (s) EID Estimated Map Fig. 1. Example stiffness reconstruction during ergodic exploration. The trajectory concentrates on the stiff inclusion while ensuring exploration of the rest of the surface. The segmentation step extracts the inclusion boundary. F/T Sensor Indenter Ur3e Elasticity Map Ground-Truth Fig. 2. Experimental setup: a UR3e robot palpates a silicone phantom with a spherical indenter and a force/torque sensor. the correct number of inclusions and reconstructed their shapes. The segmentation step achieved high sensitivity and IoU, confirming that both localisation and boundary delineation were accurate. The trajectory generated can be seen in Figure 1. B. Phantom Validation A silicone phantom with an embedded stiff spherical inclusion was fabricated to mimic biological tissue, Figure 2, using the Ecoflex-0030 for the softer matrix and the DragonSkin30NV for the inclusion. The robot used was a Ur3e equipped with a BOTA-SensOne F/T sensor. In all trials, the inclusion was correctly detected and segmented; in Figure 3, the trajectory on the silicone sample with the estimated elasticity map is shown. The reconstructed maps closely matched the ground truth, with a root mean square error (RMSE) of 10 MPa for the elasticity map, and segmentation results showing a sensitivity above 0.9 and a specificity above 0.95 across five experiments. Despite sensor noise and material variability, the framework achieved robust and consistent performance. IV. CONCLUSION This work presents an autonomous palpation framework that integrates viscoelastic estimation, probabilistic mapping, er0 10 20 30 40 50 x (mm) 0 10 20 30 40 50 y (mm) Estimated Elastic modulus 102 111 120 129 138 147 156 165 174 183 Elastic modulus (GPa) Fig. 3. Palpation trajectory (colour-coded from black to white, indicating progression over time) overlaid on the reconstructed elasticity map. godic trajectory planning, and segmentation. By continuously balancing exploration and refinement, the robot can reconstruct stiffness maps and delineate inclusions without requiring specialised tactile sensors. Simulation and phantom experiments confirm that the method achieves robust detection and accurate segmentation of stiff inclusions, even under noise. Future work will extend the approach to three-dimensional palpation on curved surfaces and validate it on anatomically realistic phantoms and ex-vivo tissues. Ultimately, ergodic palpation could provide clinicians with an objective, automated, and reliable diagnostic tool, complementing medical imaging and enhancing robot-assisted interventions. REFERENCES [1] E. Ayvali, A. Ansari, et al., “Utility-guided palpation for locating tissue abnormalities,” IEEE Robotics and Automation Letters, vol. 2, no. 2, 2017. [2] S. Ivi´ c, B. Crnkovi´ c, et al., “Ergodicity-based cooperative multiagent area coverage via a potential field,“ IEEE Trans. on Cybernetics, vol. 47, 2017. [3] L. Beber, E. Lamon, et al., “Force-Based Viscosity and Elasticity Measurements for Material Biomechanical Characterization With a Collaborative Robotic Arm,” IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–14, 2025. 98