💻 computer science

Autonomous Robotic Tissue Palpation and Abnormalities Characterisation via Ergodic Exploration

This paper presents an autonomous robotic palpation framework that integrates viscoelastic modeling, ergodic exploration driven by Expected Information Density, and adaptive trajectory planning to achieve superior real-time elastic mapping and abnormality detection compared to existing Bayesian optimization techniques.

Original authors: Luca Beber, Edoardo Lamon, Matteo Saveriano, Daniele Fontanelli, Luigi Palopoli

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Luca Beber, Edoardo Lamon, Matteo Saveriano, Daniele Fontanelli, Luigi Palopoli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to find a small, hard lump inside a soft, squishy piece of dough (like a tumor in breast tissue). In the real world, doctors do this by palpation—feeling the tissue with their fingers. They press down, feel how much it resists, and map out where the hard spots are.

But human hands have limits. They get tired, they might miss small lumps, and every doctor feels things slightly differently.

This paper proposes a robotic doctor's hand that is smarter, faster, and more consistent. Here is how it works, broken down into simple concepts:

1. The Robot's "Finger" and "Brain"

Instead of a fancy, expensive sensor built into the robot's finger, the team uses a standard force sensor (like a very sensitive scale) attached to a robotic arm.

The Analogy: Think of the robot's finger as a blind person's cane. It doesn't "see" the tissue; it "feels" the resistance.
The Magic: As the robot presses and moves, it uses a special math model (a "viscoelastic model") to figure out exactly how stiff the tissue is at that exact spot. It's like the robot is asking, "If I push this hard, how much does it squish back?" and instantly calculating the stiffness.

2. The Problem: How to Search?

If you were looking for a needle in a haystack, you could:

Method A (The Old Way - Bayesian Optimization): Pick a spot, check it. If it looks promising, check nearby spots. If it looks boring, jump far away. This is like a detective following a hunch. It's good at finding the center of the needle, but it often misses the edges or the shape of the needle.
Method B (The New Way - Ergodic Exploration): Imagine you are a gardener watering a lawn. You don't just water the spots you think are dry; you move your hose in a smooth, flowing pattern that covers the whole lawn, but you spend more time watering the dry patches and less time on the wet ones.

This paper uses Method B. It calls this Ergodic Exploration.

3. The "Heat Map" Guide

The robot doesn't just wander randomly. It has a mental "Heat Map" (called Expected Information Density) that tells it where to go.

The Metaphor: Imagine the tissue is a dark room. The robot wants to turn on the lights.
- Uncertainty: If the robot hasn't touched a spot yet, that spot is "dark" (high uncertainty). The robot wants to go there to turn on the light.
- Stiffness: If the robot feels something hard, that spot is "hot." The robot wants to linger there to see how big the hard spot is.
- Edges: If the hardness changes quickly (like the edge of a tumor), that's a "bright line." The robot wants to trace that line carefully.

The robot's brain combines all these feelings into a single map and moves its hand in a smooth, continuous flow, spending more time in the "interesting" areas and less time in the "boring" soft areas.

4. The "Heat Equation" Controller

How does the robot move so smoothly? It uses a mathematical trick called the Heat Equation Driven Area Coverage (HEDAC).

The Analogy: Think of dropping a drop of ink into a glass of water. The ink spreads out smoothly, filling the gaps. The robot's path is like that spreading ink. It naturally flows into the gaps it hasn't visited yet and smooths out its path to avoid jerky, sudden jumps. This allows the robot to move at a steady, safe speed (100 times a second!) without getting stuck or missing spots.

5. When Does It Stop?

Usually, robots stop after a set time (e.g., "search for 5 minutes"). This paper uses a smarter rule: The robot stops when it feels like it has "seen" enough.

The Metaphor: Imagine you are reading a book. You don't stop after 10 pages; you stop when you understand the story. The robot measures how well its map matches the "ideal" coverage. Once the map is detailed enough and the robot isn't finding any new surprises, it says, "Okay, I'm done," and stops. This saves time and battery.

6. The Results: Did It Work?

The team tested this in two ways:

Computer Simulations: They created fake "tumors" in a digital world. The robot found them faster and drew a more accurate picture of their shape than the old "detective" method (Bayesian Optimization).
Real-World Test: They built a fake tissue out of soft silicone with a hard ball hidden inside (like a tumor). The robot successfully found the ball, mapped its shape, and distinguished it from the soft tissue, even when there was "noise" (like shaky hands or sensor errors).

Why Does This Matter?

It's Cheaper: It uses standard robot parts, not expensive custom sensors.
It's Safer: It moves smoothly, not in jerky jumps, which is crucial for delicate medical procedures.
It's Smarter: It knows when to explore new areas and when to focus on finding the details of a problem.

In a nutshell: This paper teaches a robot how to "feel" its way around soft tissue, not by randomly poking, but by flowing like water over a landscape, spending extra time investigating the bumps and hard spots, and stopping only when it has a complete and accurate picture of what's underneath.

1. Problem Statement

Robotic palpation is critical for early diagnosis (e.g., breast cancer) and safe navigation in robot-assisted minimally invasive surgery (RMIS). However, existing approaches face several limitations:

Subjectivity & Variability: Manual palpation is subjective and prone to inter-practitioner variability.
Hardware Constraints: Many existing robotic palpation systems rely on custom, embedded tactile sensors or expensive imaging (MRI/Ultrasound) that are difficult to integrate into standard surgical robots.
Algorithmic Limitations: Previous autonomous methods often use Bayesian Optimisation (BO). While BO is effective at finding local maxima (stiff peaks), it struggles to accurately delineate the boundaries of abnormalities and can be inefficient at balancing global exploration with local exploitation. Furthermore, standard BO generates discrete, point-to-point sampling paths that are unsafe or impractical for continuous robotic motion in clinical settings.

Goal: Develop an autonomous framework that uses only a standard commercial Force/Torque (F/T) sensor to create real-time, high-fidelity elastic stiffness maps and detect stiff inclusions (tumors) without specialized tactile hardware.

2. Methodology

The proposed framework is a closed-loop system integrating online viscoelastic parameter estimation with ergodic trajectory planning.

A. Sensing and Parameter Estimation

Hardware: Uses a standard 6-axis F/T sensor (Bota SensOne) and a spherical indenter attached to a UR3e robot.
Model: Employs a viscoelastic contact model based on the Dimensionality Reduction Method (DRM). This models tissue response as a combination of Hertzian elastic deformation and viscoelastic damping.
Estimation: An Extended Kalman Filter (EKF) estimates the tissue's Young's modulus (stiffness) and viscosity in real-time from force and indentation depth/velocity data. This allows for continuous mapping without stopping for discrete measurements.

B. Target Distribution: Expected Information Density (EID)

Instead of a simple uncertainty map, the authors define a tailored Expected Information Density (EID) function, $\xi_{EID}(x)$ , to guide exploration. This function balances three factors:

Uncertainty: High variance in the Gaussian Process Regression (GPR) model (exploration).
Stiffness Magnitude: High estimated stiffness values (exploitation of potential lesions).
Spatial Gradients: High gradients in stiffness (refining boundaries of lesions).

The trade-off is controlled by a time-varying scalar $\alpha(t)$ , derived from the ergodic metric. Initially, $\alpha$ is high, prioritizing exploration of uncertain areas. As the robot covers the space, $\alpha$ decreases, shifting focus toward exploiting high-stiffness regions and refining boundaries.

C. Trajectory Planning: Ergodic Control via HEDAC

Algorithm: The system uses Heat Equation Driven Area Coverage (HEDAC), a real-time ergodic controller.
Mechanism: Unlike Spectral Multiscale Coverage (SMC), which is computationally heavy and limited to low frequencies (~4 Hz), HEDAC solves a stationary heat equation to generate smooth, continuous trajectories.
Performance: The controller runs at 100 Hz, ensuring smooth motion and rapid adaptation to new data. It generates trajectories where the time spent in any region is proportional to the EID of that region.

D. Stopping Criterion

The system employs an information-theoretic stopping criterion based on the ergodic metric $\alpha$ . Exploration terminates when $\alpha$ reaches a threshold (e.g., 0.4), indicating that the trajectory has sufficiently converged to the target distribution, avoiding arbitrary time limits or fixed step counts.

3. Key Contributions

Tailored EID for Stiffness Mapping: A novel information density function that explicitly combines uncertainty, stiffness magnitude, and spatial gradients to guide ergodic exploration toward diagnostically relevant regions.
Closed-Loop Integration: A seamless loop between online force-based viscoelastic estimation and ergodic trajectory planning, where stiffness estimates continuously reshape the exploration objective.
Real-Time Implementation: Successful deployment of ergodic palpation using HEDAC at 100 Hz, enabling continuous motion and trade-offs between exploration and exploitation, overcoming the computational bottlenecks of previous spectral methods.
Hardware Agnosticism: The framework relies solely on a commercial F/T sensor, eliminating the need for custom tactile arrays and ensuring portability and cost-effectiveness.

4. Results

Simulation Results

Setup: Tested on synthetic stiffness maps with 1, 2, and 3 stiff inclusions (simulating tumors) in a soft background.
Comparison: The proposed Ergodic method was compared against two Bayesian Optimisation baselines: BO-EI (standard Expected Improvement) and BO-EIS (BO with intermediate sampling for continuous motion).
Performance:
- Accuracy: The Ergodic approach achieved significantly lower Root Mean Square Error (RMSE) in stiffness reconstruction compared to both BO variants (e.g., 1.42 kPa vs. 1.87 kPa for BO-EIS at 800mm travel in noise-free conditions).
- Efficiency: It reached high detection rates (99%) with shorter execution times and less trajectory length than BO.
- Segmentation: In boundary delineation tasks (sensitivity/specificity), the Ergodic method outperformed BO, particularly in detecting irregular shapes (Sensitivity ~0.975 vs. 0.950 for BO-EIS).
- Robustness: The method maintained high performance even with added measurement noise ( $N(0, 6.25 \text{ kPa}^2)$ ).

Experimental Validation

Setup: A UR3e robot palpated a silicone phantom containing a stiff spherical inclusion (DragonSkin-30NV) embedded in a soft matrix (Ecoflex-0030).
Outcomes:
- Successfully identified the location of the stiff inclusion in all 5 trials.
- Reconstruction Accuracy: RMSE of $9.76 \pm 2.97$ kPa.
- Classification: Average Sensitivity of 0.92 and Specificity of 0.98, demonstrating a strong ability to avoid false negatives (critical for medical screening).
- The trajectory naturally concentrated on the stiff region while covering the rest of the map, validating the adaptive nature of the EID.

5. Significance and Conclusion

This work demonstrates that ergodic exploration is superior to Bayesian Optimisation for continuous robotic tissue palpation. By shifting from discrete point-sampling to continuous, density-proportional trajectory planning, the system achieves:

Higher Fidelity Maps: More accurate reconstruction of stiffness boundaries, which is vital for defining tumor margins.
Clinical Viability: The use of off-the-shelf F/T sensors makes the system compatible with standard surgical robots, lowering the barrier to adoption.
Safety and Efficiency: The continuous, smooth trajectories generated by HEDAC are safer for patient interaction than the erratic jumps often associated with BO.

The proposed framework offers a robust, real-time solution for autonomous diagnostic palpation, capable of detecting and characterizing tissue abnormalities with high sensitivity and specificity, paving the way for assistive diagnostic tools in clinical and surgical settings. Future work aims to extend this to 3D non-flat surfaces and ex vivo tissues.