Sensorless strain-bounded grasping of fruit

arXiv:2608.23983 · cs.RO · Submitted 2026-08-25 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Sensorless strain-bounded grasping of fruit".

Dev: Here's the gist of the paper. Robotic fruit harvesting must hold produce securely without bruising it, yet compression stiffness varies several-fold with ripeness within a single species, so no fixed grip force spans the range <ref:2608.23983#pg2> The controller closes the gripper until the object’s estimated compression strain reaches a user-specified limit epsilon, using only the encoder position and motor-effort signal on every servo gripper—no tactile or force-torque sensor <ref:2608.23983#pg2> Dividing an effort-based contact force by a lower bound on object stiffness makes the stop provably conservative—true compression stays at or below epsilon—for any epsilon above a contact-detection strain floor we identify and quantify <ref:2608.23983#pg2> Core Methodology The controller consumes two scalar signals per control tick at 50 Hz: the gripper jaw position x(t) (encoder) and a motor-effort proxy I(t) (current in simulation; the loadregister reading — a duty-cycle proxy — on the Feetech servo in hardware) <ref:2608.23983#pg2> It produces a single normalized action a(t) ∈

−1, 1: that commands jaw velocity <ref:2608.23983#pg2> From these two signals alone, with no tactile sensor and no force-torque measurement, it estimates the object’s compression strain ̂r(t) — the fraction by which the object is compressed relative to its size — in real time <ref:2608.23983#pg2> This strain is recovered from an estimated contact force and a fixed stiffness lower bound, ̂r = F̂/(kminD), not from jaw travel alone <ref:2608.23983#pg2> Certification of Safety The controller divides the force proxy by the stiffness lower bound kmin, and this bound certifies a damage-safe stop <ref:2608.23983#pg2> The compression strain r = F/(kD) is decreasing in stiffness, and every admissible object satisfies ktrue ≥ kmin, so the estimate over-bounds the truth (Fig.

Rosa: 3) <ref:2608.23983#pg2> Stopping when ̂r = epsilon therefore caps the true compression at rtrue = epsilonkmin/ktrue ≤ epsilon, with equality only for the softest object <ref:2608.23983#pg2> The user bound epsilon is thus a certified upper limit on the true deformation <ref:2608.23983#pg2> Performance in Validation In MuJoCo simulation over a realistic fruit-stiffness range, under a sensor-noise model calibrated to the real servo, the controller holds ≥ 98 % grasp at 0 % damage across all medium-to-firm stiffnesses for the entire certified epsilon range, which neither fixed-force baseline attains <ref:2608.23983#pg2> On stiffness-graded 3D-printed TPU cubes it matches baseline grasp success at roughly half the grip force and cuts soft-object damage from 100 % to 40 % <ref:2608.23983#pg2> Operational Control Knob The closing speed v is a fourth, operational setting — a throughput–gentleness tradeoff whose effect on the smallest attainable epsilon we quantify in §6.3 <ref:2608.23983#pg2> The detection floor is linear in v with the threshold term as its intercept (Fig.

Dev: 7) <ref:2608.23983#pg2> Deployments should choose v so that epsilonf loor(v, kmax) sits below the smallest epsilon the task requires <ref:2608.23983#pg2> Limitations and Future Directions The very-soft regime is where the single dial must trade off <ref:2608.23983#pg2> Real-fruit validation (e.g., on tomato) is therefore the most important next step <ref:2608.23983#pg2> A natural extension is to relax the fixed-D input by estimating an effective contact dimension online <ref:2608.23983#pg2> The estimation needs excitation that the task is built to suppress <ref:2608.23983#pg2> We presented a compression-strain controller for damage-bounded grasping that uses only encoder position and motoreffort signals — no tactile or force-torque sensor <ref:2608.23983#pg2> It turns an effort-based force proxy into a compression estimate by dividing by a known lower bound on object stiffness, and a single proportional law closes the gripper until this strain reaches a user-specified bound epsilon <ref:2608.23983#pg2> Because the operative stiffness is a lower bound, the stop is provably conservative—the true compression stays at or below epsilon—for any epsilon above a contact-detection strain floor that we identify and quantify <ref:2608.23983#pg2> In simulation under a sensor-noise model calibrated to the real servo, the method holds ≥ 98 % grasp at 0 % damage at every stiffness from 4000 N∕m up, for every epsilon ∈

0.7, 2.0: % — an operating regime neither fixed-force baseline attains <ref:2608.23983#pg2> On real hardware, it achieves equivalent grasp success at medium and stiff cubes while applying roughly half the grip force of either baseline, and lowers soft-cube damage from 100 % (both baselines) to 40 % <ref:2608.23983#pg2> The same stiffness margin absorbs motor-effort calibration error, so the method ports across hardware without per-object force calibration <ref:2608.23983#pg2> its main limitation is the very-soft regime, where a conservative damage bound and a reliable lift become mutually exclusive for rigid pads <ref:2608.23983#pg2> The validation on real fruit, relaxing the fixed-diameter assumption, and recovering an online stiffness estimate — to tighten the bound on firm fruit and serve as a ripeness cue — are natural next steps <ref:2608.

Title and authors: Rosa: So we're talking about this paper, "Sensorless strain-bounded grasping of fruit." It sounds like they've figured out how to hold produce without bruising it by only looking at the position and effort signals from the gripper, no touch sensors needed.

Dev: Exactly. The core idea is that fruit stiffness changes a lot depending on ripeness—sometimes it’s super soft, sometimes firm—so you can't just use one fixed grip force for everything.

Taro: I'm curious how they handle that stiffness variation without actually having a force sensor to tell what the object feels like.

Rosa: Well, the paper explains that instead of tuning a force threshold, they bound the deformation instead. They close the gripper until an estimated compression strain hits a limit you set, say epsilon.

Dev: And they get that strain estimate from just two things: the encoder position and the motor effort signal. They use those to calculate an estimate of how compressed the object is relative to its size.

Taro: So what’s the catch there? If they aren't touching it, how do you know what force or compression is happening?

Rosa: That's where they introduce a key idea: they divide that effort-based contact force proxy by a lower bound on object stiffness. They use this lower bound to make sure their stopping condition is provably conservative.

Dev: That means the true compression stays at or below your limit epsilon, as long as your limit is above some kind of detection floor they can identify.

Taro: So they've created a certified safety boundary based on knowing the softest possible object in that ripeness range, which is what sets that lower bound.

Rosa: Right. And they quantify this contact-detection strain floor, which also shows how closing speed affects it linearly, making speed a throughput gentleness knob for the operator.

Dev: It’s interesting because this whole system relies on you setting three things before you grasp: the damage bound epsilon, the object diameter D, and that stiffness lower bound kmin.

Taro: So those three inputs—epsilon, diameter D, and kmin—are what define how the controller estimates strain in real time.

Rosa: Right. And they emphasize that epsilon is just a task specification you set; D and kmin are priors they estimate based on the object class you're holding.

Title and authors: Dev: The paper shows simulation and hardware validation across different stiffnesses, demonstrating that this method achieves lower damage than fixed-force grasping at comparable grasp success rates.

Taro: That’s a big step for real world testing because it moves away from just lab simulations and shows it works under noise models calibrated to actual servo motors.

Rosa: And on the TPU cubes they tested, they matched the baseline grasp success while using about half the grip force and cutting soft-object damage from one hundred percent down to forty percent <ref:2608.23983#pg1>.

Dev: That stiffness margin absorbing motor effort calibration error is a nice touch because it means you don't need to calibrate per object; you can just use this single method across different hardware without re-tuning everything for each new fruit.

Taro: But they also admit a limitation—the very soft regime is where things get tricky, and the authors say that in that range, the conservative damage bound and a reliable lift might start to conflict.

Rosa: So the paper points toward real fruit validation as the next most important step because those are where you'll really test that soft regime trade-off.

Dev: If you look at how they define closing speed as a setting, it directly impacts the smallest strain epsilon you can actually achieve, which is quantified in section six point three.

Taro: It sounds like operators will have to choose a closing speed that keeps their desired epsilon above the minimum attainable value for that specific object stiffness and diameter combination.

Rosa: Exactly. So the next thing we're looking at is how this method scales when you move beyond simple cubes and into more complex, real-world fruit shapes.

Dev: And what we do next is really figure out how robust this whole setup is when the world gets messy, which brings us to the improvements they suggested for their work.

Taro: What are these proposed improvements, specifically? They seem focused on making it even more practical outside of just lab tests.

Rosa: Well, they suggest relaxing that fixed diameter input D by estimating an effective contact dimension online instead of requiring you to know it beforehand.

Dev: That would mean the system has to learn how big the object is while it's closing, which adds a lot more complexity to the control loop and requires good excitation signals.

Title and authors: Taro: I agree, that sounds like something that needs careful design because if the estimation relies on what it's trying to control, you need a very stable setup.

Rosa: And they also touch on how this whole method inherits a pre-certified safety parameter when you train an AI policy for harvesting—meaning the policy just gets this built-in guarantee automatically.

Dev: That's the key for deployment: giving an AI agent a starting point where it already knows its damage constraints are managed safely above the detection floor.

Taro: So they're trying to bake that conservative stopping behavior directly into how you train the policy, which is a smart way to handle real-world failures.

Rosa: Overall, this paper on sensorless strain-bounded grasping of fruit shows a new way to grasp produce using only position and effort signals without needing tactile sensors.

Dev: It turns an effort-based force proxy into a compression estimate by dividing it by the known lower bound on object stiffness, making the stop provably conservative.

Taro: The results in simulation and hardware validation show it holds over ninety-eight percent grasp success at zero damage across medium to firm stiffnesses for the entire certified epsilon range of zero point seven to two point zero percent.

Rosa: On real hardware, they saw equivalent grasp success on stiffness-graded three dimensional printed tpu cubes while applying roughly half the grip force of either baseline and lowering soft object damage from one hundred percent down to forty percent.

Dev: The method's mechanism ensures that the stiffness margin absorbs motor effort calibration error up to the stiffness ratio, which means it requires no per object force calibration for deployment.

Taro: The main limitation they flag is that in the very soft regime, you have to trade off between a conservative damage bound and a reliable lift because those two things can become mutually exclusive for rigid pads.

Rosa: So the path forward seems clear: real fruit validation on something like tomatoes and developing an online stiffness estimate to tighten that bound on firm fruit.

Dev: That's what we need to focus on next, moving from theoretical bounds in simulation to robust performance when the world misbehaves and misbehaves in unexpected ways.

The paper's summary: Rosa: So, to wrap up what we just heard, this paper is about getting you to hold fruit without touching it at all by just looking at where your gripper is and how hard you're pushing—the motor effort.

Dev: Right. It’s building a whole system that estimates the compression strain of the fruit in real time, using only those two signals, no touch sensors, no force-torque sensors.

Taro: So it’s essentially turning an effort signal into a compression estimate by dividing it by a known lower bound on how stiff the object can possibly be.

Rosa: Exactly. That lower bound is what makes the whole thing safe because it’s provably conservative; the actual compression stays below your limit as long as you set your limit above some baseline floor they can detect.

Dev: And that means if you set a damage limit, epsilon, anything above that floor is actually damage-safe for any object in that stiffness range.

Taro: It’s like they’ve quantified the softest possible object and used it to set a safety zone around your desired grip force.

Rosa: That’s right. They showed in simulation, even under noise models that match real robot servos, this method works at over ninety-eight percent grasp success with zero damage across a whole range of stiffnesses for an epsilon between zero point seven and two point zero percent.

Dev: And on the physical hardware they tested—those three dee printed TPU cubes—it matched the baseline grasp success but used about half the grip force, and it cut soft-object damage from one hundred percent down to forty percent.

Taro: That means if you’re harvesting something that might be soft, this system is much better at not bruising it than just using a fixed force setting.

Rosa: It also handles the calibration errors in your robot's motor effort signal because the stiffness margin absorbs that error; you don't need to recalibrate for every single fruit.

Dev: But there’s a catch, and this is where we need to watch things closely. They admit that the very soft regime, where you want to be extra gentle, is where the conservative damage bound and a reliable lift might start conflicting with each other.

Taro: Yeah, that sounds like a trade-off problem; trying to be super careful about damage often means you can't get a secure hold on something really squishy.

Rosa: That’s the limitation they flag. So, for now, it seems best to stick with this method for medium and firm objects where it’s doing such a good job reducing damage compared to traditional methods.

Dev: The implication here is that we can design grasping policies that are inherently safer against damage without needing complex, expensive tactile sensors on every end effector.

Taro: It moves the safety requirement from "don't touch" to "don't exceed this calculated strain," which is a different kind of constraint for training a robot.

Rosa: And looking ahead, they suggest that next steps involve using an online stiffness estimate instead of assuming a fixed object size, which would make it even more adaptable to real fruit.

The paper's improvements: Rosa: So, to finish up, the paper isn't just stopping there; they’re suggesting ways to make this system even better for real-world use on fruit.

Dev: Right, they talk about relaxing that fixed diameter input D—meaning you don't have to know exactly how big the object is before you start grasping it.

Taro: If you estimate the contact dimension online, that means the robot has to learn about its size while it’s already trying to hold onto something.

Rosa: That’s right. It implies a more dynamic system where the grasping strategy adjusts based on what it's actually interacting with, instead of relying on a pre-set guess.

Dev: And they say that this online estimation needs some kind of excitation signal to work reliably, and this excitation is something the task itself is built to suppress.

Taro: So they’re saying the robot needs a way to intentionally wiggle or move slightly to help it figure out the object's size without destroying it.

Rosa: Precisely. It sounds like we need to build that in, because if we don't have that controlled movement, the online estimation won't be accurate enough for safe grasping.

Dev: They also mentioned how this whole strain-bounded method can actually be integrated into training an AI policy for harvesting fruit.

Taro: That’s smart; it means a policy learning to harvest can inherit this pre-certified safety parameter automatically, so the robot knows its damage limits are already managed above that detection floor.

Rosa: It gives the learned behavior a built-in guarantee of safety, which is huge for deployment because you don't have to painstakingly tune every single parameter for every new fruit type.

Dev: That solves the calibration problem we talked about earlier, where you didn't need per-object force calibration because the stiffness margin handled that error.

Taro: So they’re moving toward a more robust autonomy where the safety constraints are baked into the learning process rather than being an afterthought added on top.

Rosa: It sounds like this is exactly what we need to see—a way for robots to handle uncertainty and ripeness variation gracefully while maintaining a strict damage-safe envelope.

Dev: But we still have that big limitation they flagged: the very soft regime. They admit that in that area, trying to be super safe about damage can actually conflict with trying to get a secure lift.

Conclusion: Rosa: So, to wrap up this whole thing on "Sensorless strain-bounded grasping of fruit," we're looking at a way to grasp produce using just position and effort signals without any touch sensors at all.

Dev: Right. It turns that raw motor effort into a real-time estimate of compression strain by dividing it by a known lower bound on how stiff the object can be.

Taro: And because they use that stiffness lower bound, the stopping condition is provably conservative; the true compression stays at or below your limit as long as you set your limit above some baseline floor they can detect.

Rosa: That means if you set a damage limit, epsilon, anything above that floor is actually damage-safe for any object in that stiffness range.

Dev: And they showed in simulation, even under noise models that match real robot servos, this method works at over ninety-eight percent grasp success with zero damage across a whole range of stiffnesses for an epsilon between zero point seven and two point zero percent.

Taro: And on the physical hardware they tested—those three dee printed TPU cubes—it matched the baseline grasp success but used about half the grip force, and it cut soft-object damage from one hundred percent down to forty percent.

Rosa: It also handles those calibration errors in your robot's motor effort signal because the stiffness margin absorbs that error; you don't need to recalibrate for every single fruit.

Dev: But there’s a catch, and this is where we need to watch things closely. They admit that the very soft regime, where you want to be extra gentle, is where the conservative damage bound and a reliable lift might start conflicting with each other.

Taro: Yeah, that sounds like a trade-off problem; trying to be super careful about damage often means you can't get a secure hold on something really squishy.

Rosa: So, for now, it seems best to stick with this method for medium and firm objects where it’s doing such a good job reducing damage compared to traditional methods.

Dev: The implication here is that we can design grasping policies that are inherently safer against damage without needing complex, expensive tactile sensors on every end effector.

Taro: It moves the safety requirement from "don't touch" to "don't exceed this calculated strain," which is a different kind of constraint for training a robot.

Rosa: And looking ahead, they suggest that next steps involve using an online stiffness estimate instead of assuming a fixed object size, which would make it even more adaptable to real fruit.

Dev: That’s what we need to focus on next; figuring out how this whole system scales when you move beyond simple cubes and into more complex, real-world fruit shapes.

Taro: I agree; that online estimation needs some kind of excitation signal to work reliably, and that's something the task itself is built to suppress.

Kyushu University

cs.RO

Submitted: 2026-08-25

Updated: 2026-11-12

Comments: 14 pages, 10 figures. Revised version submitted to Robotics and Autonomous Systems

Code: https://github.com/shutouyusei/damage-safe-grasping

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: Robotic fruit harvesting must hold produce securely without bruising it, yet compression stiffness varies several-fold with ripeness within a single species…

Key concepts

Compression Strain Estimation
The controller estimates how much an object is being compressed relative to its size. It achieves this by dividing an estimated contact force proxy (derived from motor effort) by a known lower bound on the object's stiffness. This allows it to estimate strain without needing direct force or tactile sensors.
Stiffness Lower Bound ($k_{min}$)
This is a minimum stiffness value identified for the objects being grasped. Because the actual stiffness of any object is guaranteed to be greater than or equal to this lower bound, using $k_{min}$ in the calculation ensures that the resulting strain estimate provides a conservative upper bound on the true deformation.
Certified Conservative Guarantee
The method provides a mathematical guarantee that if you set your stopping limit $\epsilon$ above a specific detection floor, the actual compression will never exceed $\epsilon$. This means the system is provably damage-safe for any user-specified limit within this safe range.
Throughput-Gentleness Knob ($v$)
This is an operational setting that controls the closing speed of the gripper. It allows users to trade off harvesting speed (throughput) against how gentle the grasp is. The system identifies a detection floor for this speed, helping operators choose a closing speed that respects their desired damage limit.

Terminology

Summary

Robotic fruit harvesting must hold produce securely without bruising it, yet compression stiffness varies several-fold with ripeness within a single species, so no fixed grip force spans the range <ref:2608. p. 023983#pg2> The controller closes the gripper until the object’s estimated compression strain reaches a user-specified limit epsilon, using only the encoder position and motor-effort signal on every servo gripper—no tactile or force-torque sensor <ref:2608. p. 023983#pg2> Dividing an effort-based contact force by a lower bound on object stiffness makes the stop provably conservative—true compression stays at or below epsilon—for any epsilon above a contact-detection strain floor we identify and quantify <ref:2608. p. 023983#pg2>

Core Methodology

The controller consumes two scalar signals per control tick at 50 Hz: the gripper jaw position x(t) (encoder) and a motor-effort proxy I(t) (current in simulation; the loadregister reading — a duty-cycle proxy — on the Feetech servo in hardware) <ref:2608.23983#pg2> It produces a single normalized action a(t) ∈ [−1, 1] that commands jaw velocity <ref:2608. p. 123983#pg2> From these two signals alone, with no tactile sensor and no force-torque measurement, it estimates the object’s compression strain ̂r(t) — the fraction by which the object is compressed relative to its size — in real time <ref:2608. p. 123983#pg2> This strain is recovered from an estimated contact force and a fixed stiffness lower bound, ̂r = F̂/(kminD), not from jaw travel alone <ref:2608. p. 123983#pg2>

Certification of Safety

The controller divides the force proxy by the stiffness lower bound kmin, and this bound certifies a damage-safe stop <ref:2608.23983#pg2> The compression strain r = F/(kD) is decreasing in stiffness, and every admissible object satisfies ktrue ≥ kmin, so the estimate over-bounds the truth (Fig. p. 3 3) <ref:2608. p. 023983#pg2> Stopping when ̂r = epsilon therefore caps the true compression at rtrue = epsilonkmin/ktrue ≤ epsilon, with equality only for the softest object <ref:2608. p. 323983#pg2> The user bound epsilon is thus a certified upper limit on the true deformation <ref:2608. p. 323983#pg2>

Performance in Validation

In MuJoCo simulation over a realistic fruit-stiffness range, under a sensor-noise model calibrated to the real servo, the controller holds ≥ 98 % grasp at 0 % damage across all medium-to-firm stiffnesses for the entire certified epsilon range, which neither fixed-force baseline attains <ref:2608.23983#pg2> On stiffness-graded 3D-printed TPU cubes it matches baseline grasp success at roughly half the grip force and cuts soft-object damage from 100 % to 40 % <ref:2608. p. 023983#pg2>

Operational Control Knob

The closing speed v is a fourth, operational setting — a throughput–gentleness tradeoff whose effect on the smallest attainable epsilon we quantify in §6.3 <ref:2608. p. 023983#pg2> The detection floor is linear in v with the threshold term as its intercept (Fig. p. 6 7) <ref:2608. p. 023983#pg2> Deployments should choose v so that epsilonf loor(v, kmax) sits below the smallest epsilon the task requires <ref:2608. p. 623983#pg2>

Limitations and Future Directions

The very-soft regime is where the single dial must trade off <ref:2608.23983#pg2> Real-fruit validation (e. p. 4g., on tomato) is therefore the most important next step <ref:2608. p. 823983#pg2> A natural extension is to relax the fixed-D input by estimating an effective contact dimension online <ref:2608. p. 823983#pg2> The estimation needs excitation that the task is built to suppress <ref:2608. p. 823983#pg2>

Conclusion

We presented a compression-strain controller for damage-bounded grasping that uses only encoder position and motoreffort signals — no tactile or force-torque sensor <ref:2608.23983#pg2> It turns an effort-based force proxy into a compression estimate by dividing by a known lower bound on object stiffness, and a single proportional law closes the gripper until this strain reaches a user-specified bound epsilon <ref:2608. p. 823983#pg2> Because the operative stiffness is a lower bound, the stop is provably conservative—the true compression stays at or below epsilon—for any epsilon above a contact-detection strain floor that we identify and quantify <ref:2608. p. 823983#pg2> In simulation under a sensor-noise model calibrated to the real servo, the method holds ≥ 98 % grasp at 0 % damage at every stiffness from 4000 N∕m up, for every epsilon ∈ [0. p. 57, 2.0] % — an operating regime neither fixed-force baseline attains <ref:2608. p. 823983#pg2> On real hardware, it achieves equivalent grasp success at medium and stiff cubes while applying roughly half the grip force of either baseline, and lowers soft-cube damage from 100 % (both baselines) to 40 % <ref:2608. p. 523983#pg2> The same stiffness margin absorbs motor-effort calibration error, so the method ports across hardware without per-object force calibration <ref:2608. p. 823983#pg2> its main limitation is the very-soft regime, where a conservative damage bound and a reliable lift become mutually exclusive for rigid pads <ref:2608. p. 823983#pg2> The validation on real fruit, relaxing the fixed-diameter assumption, and recovering an online stiffness estimate — to tighten the bound on firm fruit and serve as a ripeness cue — are natural next steps <ref:2608. p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Sensorless damage-safe grasping controller: The system can close a gripper until an object’s estimated compression strain reaches a user-specified limit epsilon using only encoder position and motor effort signals, avoiding tactile or force-torque sensors.

  2. Certified conservative compression strain guarantee: The controller provides a certified conservative compression-strain guarantee where the true compression stays at or below epsilon for any epsilon above a contact-detection strain floor we identify and quantify.

  3. Throughput-gentleness knob: The closing speed becomes an explicit throughput–gentleness knob, allowing operators to balance closing speed with the desired damage limit.

  4. Stiffness margin absorption: The method's mechanism ensures that the stiffness margin absorbs motor-effort calibration error up to the stiffness ratio, meaning it requires no per-object force calibration.

  5. Safe action space for learned grasping: A policy learning a harvesting task can inherit a pre-certified safety parameter, as any epsilon selected above the detection floor stays damage-safe whenever the object class is lower bounded by kmin.

Related papers