Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

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 Mazzolai

November 2026

TL;DR A soft suction cup with three LED-phototransistor pairs inside estimates the force it presses on a kiwifruit from how its light paths deform. On 70 real fruits held out from training, pressed by a UR5e, it predicts the contact force with R² = 0.968 and 0.138 N mean error.

What they do

For handling delicate fruit, the authors adapt an optoelectronic soft suction cup to kiwifruit. Then they test how well its optical signals estimate the normal contact force on real, biologically variable fruit that the model never saw in training. Firmness classification from the same signals is a secondary, exploratory test.

How it works

  • Sensing. Three LEDs and three phototransistors in the cup stalk read how cup deformation changes the internal light. LEDs alternate on and off to cancel ambient light. Signals are sampled at 200 Hz.
  • Cup design. A Dragon Skin 10 body gave the highest pull-off force on kiwifruit (4.74 N). A soft Ecoflex 00-10 rim for gentler contact lowers it to 3.14 N, still enough to lift a whole fruit by suction.
  • Data and models. A UR5e pressed the cup (without suction) onto 70 kiwifruits to setpoints of 1-3.5 N: 1,509 trials over four days, with a load cell giving ground truth. Isotonic regression, ridge regression and a small MLP map optical signals to force. They are validated with whole fruits held out.

Results

  • Held-out fruits. The MLP reaches R² = 0.968 and MAE 0.138 N. Isotonic regression on the mean signal alone reaches R² = 0.925, while a control without optical input reaches only 0.212.
  • Held-out days. R² = 0.955, MAE 0.171 N.
  • Phases. Loading is more accurate than unloading (R² 0.979 vs 0.953).
  • Firmness. Soft/firm classification reaches 0.705 balanced accuracy, three classes only 0.490.

Limitations

Force is estimated only during known contact, with the load cell defining contact intervals, and without active suction. So this is not yet autonomous sensing during a suction grasp. Error varies by fruit (0.09-0.31 N), relative error near contact onset is 18.8%, and firmness assessment is not reliable.

Figures

From Figure 1: (a) Compression shortens and elongation lengthens the light paths between LEDs and phototransistors. (...
From Figure 1: (a) Compression shortens and elongation lengthens the light paths between LEDs and phototransistors. (b) The cup: a PCB with three LED-phototransistor pairs, Ecoflex 00-50 optics, a Dragon Skin 10 body and a softer Ecoflex 00-10 rim, 23 mm across.
From Figure 4: Suction grasp of a whole kiwifruit: approach, contact with suction on, and lift.
From Figure 4: Suction grasp of a whole kiwifruit: approach, contact with suction on, and lift.
From Figure 5: A UR5e presses the sensorized cup down on the fruit while an axial load cell under the support plate m...
From Figure 5: A UR5e presses the sensorized cup down on the fruit while an axial load cell under the support plate measures the reference force; both signals are recorded on a host PC.
From Figure 8: Force estimation on held-out kiwifruit. (a) Predicted versus measured force over 102,640 contact sampl...
From Figure 8: Force estimation on held-out kiwifruit. (a) Predicted versus measured force over 102,640 contact samples. (b) Per-fruit error across 70 fruits, median 0.134 N. (c) Error grows with force: 0.088 N below 1 N, 0.150 N at 1-2.5 N, 0.201 N above 2.5 N.
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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.