Mechanical Interaction Sensing for Robotic Kiwifruit Handling with an Optoelectronic Soft Suction Cup

Romina Adhami, Ebrahim Shahabi, Alessio Mondini, Samuel M. Youssef, Cosimo Della Santina, Gualtiero Volpe, Barbara MazzolaiarXiv 2026 · arXiv

This study presents a novel approach to robotic fruit handling by integrating optoelectronic sensing directly into a soft suction cup. By combining a bicomponent material design with shallow neural networks, the researchers achieved highly accurate, real-time interaction force estimation on biologically variable kiwifruits, even when the model encountered fruits it had never seen before.

TL;DR

Researchers developed a sensorized, bicomponent soft suction cup that uses internal light-path changes to estimate contact forces on kiwifruit with high precision ($R^2 = 0.968$). The system demonstrates robust generalization across different fruits and acquisition days, providing a foundation for damage-aware robotic harvesting and sorting.

What Problem Does This Paper Address?

Automating the handling of delicate produce like kiwifruit is a "Goldilocks" problem: apply too little force, and the grasp is unstable; apply too much, and you bruise the tissue. While soft robotic end-effectors provide inherent compliance to protect the fruit, they often lack the quantitative feedback needed for precise control.

Furthermore, biological products are messy. Kiwifruits are curved, irregular, and covered in hair, making it difficult for traditional sensors to provide consistent readings. Most existing tactile sensors are tested on flat, predictable laboratory proxies. This paper bridges the gap by evaluating a sensorized suction cup's performance on real, variable fruit across multiple days and conditions.

Methods

The core of the system is an optoelectronic sensing principle. A small PCB with three LED-phototransistor pairs is embedded in the suction cup's stalk. As the cup deforms during contact, the internal light paths change, modulating the intensity captured by the phototransistors (see Fig. 1a).

![Fig.

  1. Sensorized soft suction-cup design showing the optoelectronic sensing principle and final bicomponent construction.](/figures/art_158b1dab0df3481188c7b35f149a57dd)

Key methodological highlights include:

  • Bicomponent Design: The authors screened various geometries and materials (Fig. 2). They settled on a "Large" design using Dragon Skin 10 for the structural body (to maintain suction) and a thin Ecoflex 00-10 rim for the fruit-contacting interface (to ensure a gentle, conforming seal).
  • Signal Processing: To handle ambient light, the LEDs were pulsed, and the sensor used the absolute difference between consecutive "on" and "off" states.
  • Machine Learning: The team compared three models for force estimation: Isotonic regression (a simple monotonic mapping), Ridge regression (linear), and a shallow Multilayer Perceptron (MLP).
  • Validation: Crucially, they used grouped held-out-fruit validation, meaning the models were tested on entirely different individual fruits than those used for training, ensuring the results weren't just "memorizing" specific fruit shapes.

Datasets

The researchers collected a robust dataset to ensure statistical significance:

  • 70 real kiwifruits were used in the study.
  • 1,509 trials were conducted using a UR5e robotic arm (Fig. 5).
  • 4 separate collection-day batches were used to test robustness over time.
  • Reference Labels: An external S-beam load cell provided ground-truth force data, while a Zwick/Roell universal testing machine provided reference firmness labels via standardized compression tests.

Key Insights and Ideas

  • Material Synergy: The bicomponent design was essential. While the stiffest material (DS10) provided the best suction, it was too harsh. Adding the softer Ecoflex rim (DS10+E10) balanced high attachment force (3.14 N) with the compliance needed for delicate fruit (Fig. 3).
  • Information-Rich Signals: Even a simple monotonic mapping of the mean optical signal achieved an $R^2$ of 0.925. This suggests that the physical deformation of the cup is a very strong proxy for interaction force, regardless of the fruit's specific geometry.
  • Temporal Hysteresis: The sensors were more accurate during the loading phase than the unloading phase. This is likely due to the viscoelastic properties of the silicone and the fruit tissue, which don't return to their original shape instantly (Fig. 9a).
  • Firmness vs. Force: While force estimation was highly successful, classifying fruit firmness (Soft vs. Firm) was more challenging (accuracy ≈ 0.705). This indicates that the sensor primarily "sees" the deformation of its own structure rather than the internal properties of the fruit.

![Fig.

  1. Peak attachment force on kiwifruit, comparing monolithic and bicomponent designs.](/figures/art_efadf83318574ebba3f05664cacf2b33)

Key Results

  • Force Estimation Accuracy: The shallow MLP achieved a pooled out-of-fold $R^2 = 0.968$ and a Mean Absolute Error (MAE) of 0.138 N.
  • Generalization: The Leave-One-Collection-Day-Out (LODO) validation remained strong at $R^2 = 0.955$, proving the system isn't overly sensitive to daily setup variations.
  • Per-Fruit Performance: The median per-fruit $R^2$ was an impressive 0.973, showing consistent performance across the 70-fruit cohort.

![Fig.

  1. Representative kiwifruit interaction showing the three optoelectronic sensor responses (S1–S3) vs. measured force.](/figures/art_37d956aa41be4366bf535b06b24fcf4a)

Limitations and Future Work

While the results are promising, the authors note several areas for improvement:

  • Active Suction: The force estimation was evaluated in "contact mode" without active suction. Future work will need to account for the internal pressure changes when the vacuum pump is running.
  • Firmness Grading: The current architecture is a better force sensor than a firmness grader. Improving firmness assessment might require more complex features or different loading protocols.
  • Closed-Loop Control: The next step is integrating this real-time force feedback into the robot's control loop to allow the arm to automatically adjust its grip based on the fruit's response.

Summary updated

Abstract

Robotic handling of delicate fruit benefits from compliant contact and quantitative sensing of the mechanical interaction at the end-effector. This study investigates an optoelectronic soft suction cup with a kiwifruit-specific bicomponent design for estimating interaction forces during contact with real fruit. During known-contact intervals, the optical signals accurately estimate interaction force on fruits excluded from model training: the best shallow network achieves grouped out-of-fold pooled R-squared = 0.968 +/- 0.005 (MAE = 0.138 N), with a median per-fruit held-out R-squared = 0.973 (MAE = 0.134 N). A simple monotonic mapping from the mean optical response alone achieves R-squared = 0.925 +/- 0.005, showing that substantial force-related information is already present in the optical response. Leave-one-collection-day-out validation remains strong (R-squared = 0.955 +/- 0.031), supporting robustness across acquisition sessions. As an exploratory secondary analysis, binary Soft/Firm classification from compression-test reference labels achieves moderate balanced accuracy (approximately 0.705) under grouped held-out-fruit validation, while finer three-class grading shows limited performance.
BibTeX
@misc{adhami2026mechanical,
  author = {Romina Adhami and Ebrahim Shahabi and Alessio Mondini and Samuel M. Youssef and Cosimo Della Santina and Gualtiero Volpe and Barbara Mazzolai},
  title = {{Mechanical Interaction Sensing for Robotic Kiwifruit Handling with an Optoelectronic Soft Suction Cup}},
  year = {2026},
  eprint = {2611.09468},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09468},
}
Figure