💻 computer science

Sensorless damage-safe grasping

This paper presents a sensorless grasping controller that ensures damage-safe robotic fruit harvesting by bounding object deformation to a user-specified strain limit using only encoder and motor-effort data, thereby eliminating the need for tactile sensors while outperforming fixed-force baselines in both grasp success and damage reduction across varying stiffness levels.

Original authors: Yusei Shuto, Danilo Vasconcellos Vargas

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yusei Shuto, Danilo Vasconcellos Vargas

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 a robot arm reaching into a garden to pick a ripe tomato. The task seems simple, but it hides a difficult paradox. To carry the fruit without dropping it, the robot must squeeze hard enough to hold on. But to avoid bruising the delicate skin, it must squeeze gently enough not to crush it. The problem is that tomatoes are not all the same. A firm, unripe tomato is stiff and can take a strong grip, while a soft, overripe one is fragile and yields easily. A robot programmed with a single, fixed amount of squeezing force will inevitably fail: if the force is strong enough to hold the firm fruit, it will destroy the soft one; if it is gentle enough for the soft fruit, it will drop the firm one. For years, engineers have tried to solve this by adding expensive, fragile sensors to the robot's fingers to feel the pressure, but these add cost and complexity that make them impractical for real-world farming.

A team of researchers at Kyushu University in Japan has found a different path. Instead of trying to measure the exact force or feel the fruit's surface, they taught the robot to limit how much the fruit squishes. Their method relies on signals the robot already has: the position of its fingers and the electrical effort the motor is using to move them. By watching how hard the motor works as the fingers close, the system can estimate how much the fruit is being compressed relative to its size. The robot is programmed to stop closing the moment this estimated squish reaches a safe limit set by the human operator. This approach does not need special sensors or a unique setting for every single fruit. It works by ensuring the fruit is never squeezed beyond a specific percentage of its original size, regardless of whether the fruit is hard or soft.

The researchers tested this idea in two ways. First, they ran thousands of simulations on a computer, modeling a robot arm picking up cubes that represented fruit with different levels of stiffness, ranging from very soft to very firm. They compared their new method against two older, simpler strategies: one that stops the moment it touches the object, and another that closes with a fixed, strong grip. The results showed that the traditional methods struggled. The gentle, touch-stop strategy often failed to hold firmer objects, while the strong, fixed-grip strategy crushed the soft ones. In contrast, the new method successfully held the fruit without causing damage in nearly every case for objects that were medium to firm. Even on the softest objects, where a perfect balance is difficult, the new method caused significantly less damage than the old ways, reducing the rate of crushed items from total failure down to a manageable level.

To prove this worked in the real world, the team built a physical robot using standard parts and tested it on 3D-printed cubes made of a soft, rubber-like material. These cubes were printed with different internal structures to mimic the varying stiffness of real fruit. The results mirrored the computer simulations. The new controller held the cubes securely without damaging them, while the older methods either dropped the firm cubes or completely crushed the soft ones. Remarkably, the new method achieved this success using roughly half the gripping force of the traditional strong-grip approach. The robot was able to stop precisely when the material reached the safe limit, whereas the older methods kept squeezing, leading to over-compression and damage.

The key to this success is a concept the researchers call a "certified bound." Because the system assumes the fruit is at its softest possible state, it calculates a stopping point that is guaranteed to be safe for any fruit that is firmer than that assumption. If the fruit is actually harder, the robot will stop even earlier than necessary, which is safe. If the fruit is as soft as the assumption, it stops exactly at the limit. This creates a safety margin that does not require the robot to know the exact stiffness of the fruit beforehand. The only inputs needed are the size of the fruit and a chosen limit for how much it can be squished. The researchers also discovered that the speed at which the robot closes its fingers matters. Moving too fast means the robot spends a tiny bit of time closing after it first feels the touch but before it can react, which uses up some of the safe compression budget. By adjusting the closing speed, an operator can trade off between how fast the robot works and how gently it handles the fruit.

This work suggests that robots do not need complex, expensive sensors to handle delicate objects. By using the data the robot already generates and focusing on limiting deformation rather than force, it is possible to create a grasping system that is both safe and adaptable. While the current tests used simple shapes and synthetic materials, the principles offer a promising foundation for future robots that can harvest real fruit in the field. The method handles the wide variety of ripeness found in nature without needing to be reprogrammed for every single item, offering a practical solution to a problem that has long hindered the automation of agriculture.

Technical Summary: Sensorless Damage-Safe Grasping

Problem Statement
Robotic fruit harvesting faces a fundamental conflict between grasp success (holding the fruit securely) and fruit preservation (avoiding bruising). The compression stiffness of a single fruit species can vary several-fold depending on ripeness (e.g., tomato stiffness ranges from ~2,000 N/m for overripe to ~10,000 N/m for firm). Consequently, a fixed grip force cannot span this range: a force sufficient for firm fruit crushes soft fruit, while a force safe for soft fruit fails to hold firm fruit. Existing adaptive solutions often rely on dedicated tactile skins (e.g., GelSight) or vision-tactile fusion, which introduce cost, fragility, and integration complexity unsuitable for field-deployed harvesters. Furthermore, purely sensorless force estimation methods typically recover force or deformation estimates but lack a certified, user-interpretable damage bound that remains calibration-free across objects of unknown stiffness.

Methodology
The authors propose a proprioceptive deformation-bounded grasp controller that regulates grasp compression to a user-specified strain limit $\varepsilon$ using only encoder position and motor-effort signals (current or load-register readings). The system operates at 50 Hz and follows a two-stage estimation pipeline:

Contact Detection: The controller monitors the motor-effort signal $I(t)$ . Contact is declared when the signal exceeds a baseline $I_0$ by a threshold $\Delta I_{th}$ (set to 3–4 $\sigma$ of free-closing noise) for a persistence window ( $N_{persist}=3$ ticks). This amplitude-based detection avoids the unreliability of rate-based detection on soft objects.
Strain Estimation: Upon contact, the controller estimates the contact force $\hat{F}$ using a linear force-current relationship. Instead of estimating the specific stiffness of the individual fruit, the controller divides this force proxy by a known lower bound on object stiffness ( $k_{min}$ ), representing the softest produce in the target ripeness range. The compression strain $\hat{r}$ is calculated as:
$\hat{r}(t) = \frac{\hat{F}(t)}{k_{min} D}$
where $D$ is the object diameter (a per-object-class prior).

Control Logic
A single proportional control law drives the gripper velocity $a(t)$ based on the error between the estimated strain $\hat{r}$ and the user-specified bound $\varepsilon$ :
$a(t) = \text{clip}(k_p(\varepsilon - \hat{r}(t)), -v_{release}, v_{close})$
The gripper closes at full velocity until $\hat{r}$ approaches $\varepsilon$ , then decelerates to a hold. Crucially, the controller does not release if the strain drops (e.g., due to arm joint yield during lifting); it re-closes to maintain the strain at $\varepsilon$ .

Theoretical Guarantee (Certified Bound)
The method provides a provably conservative guarantee. Since the true stiffness $k_{true} \ge k_{min}$ , the estimated strain $\hat{r}$ over-bounds the true strain $r_{true}$ :
$r_{true} = \frac{F}{k_{true} D} \le \frac{F}{k_{min} D} = \hat{r}$
Therefore, stopping when $\hat{r} = \varepsilon$ ensures that the true compression $r_{true} \le \varepsilon$ . This bound holds without per-fruit calibration and is robust to errors in the motor-effort coefficient, provided the stiffness margin is sufficient.

Key Contributions

Proprioceptive Controller: A grasp controller that achieves adaptive, damage-safe grasping using only encoder and motor-effort signals, eliminating the need for tactile or force-torque sensors.
Certified Conservative Guarantee: A mathematical proof that the true compression strain stays at or below the user-specified bound $\varepsilon$ for any object with stiffness $k \ge k_{min}$ , without per-object calibration.
Quantified Detection Floor: The identification of a "contact-detection strain floor." Robust contact detection requires a finite amount of compression to cross the noise threshold and persist. This floor, which scales linearly with closing speed, defines the minimum attainable $\varepsilon$ and establishes closing speed as an explicit throughput–gentleness trade-off knob.
Validation: Comprehensive simulation and hardware validation demonstrating performance across varying stiffnesses.

Results

Simulation (MuJoCo): Using a sensor-noise model calibrated to real hardware, the proposed method achieved $\ge$ 98% grasp success with 0% damage across all medium-to-firm stiffnesses ( $k \ge 4000$ N/m) for the entire certified $\varepsilon$ range ( $0.7\% - 2.0\%$ ). Neither fixed-force baseline (gentle or aggressive) achieved this "clean corner" performance.
Hardware (3D-Printed TPU Cubes): On soft cubes, the proposed method reduced damage from 100% (both baselines) to 40% (at $\varepsilon=0.7\%$ ) while maintaining grasp success. On medium and stiff cubes, it matched baseline grasp success (100%) while applying roughly half the grip force.
Fixed-Force Baselines: Baselines failed on soft objects (100% damage or timeout) and were either too weak (gentle) or too destructive (aggressive) for the full stiffness range.

Significance and Claims
The paper claims that this approach occupies a previously empty region in the design space: a certified deformation bound derived solely from proprioceptive signals. Unlike hand-tuned force thresholds, the parameter $\varepsilon$ is:

Operator-interpretable: It represents a dimensionless strain limit applicable across different fruit sizes and robot scales.
Size-scaling: The effective force threshold automatically scales with object diameter ( $D$ ).
Safe for Learning: It provides a pre-certified safe action space for learned grasping policies, allowing exploration over $\varepsilon$ without risking fruit damage.

The authors acknowledge limitations, specifically that in the very-soft regime (where a stable lift requires compression near the damage line), the rigid-pad approach must trade grasp success for damage avoidance. They note that real-fruit validation (addressing anisotropy and irregular geometry) and online stiffness estimation (to tighten bounds on firm fruit) are necessary future steps.

Technical Summary: Sensorless Damage-Safe Grasping

More like this