Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Pick Up a Single Grape
Transcript
- Lucas: We've spent 82 episodes cataloging things robot arms still can't do. Wet towels, raw eggs, fitted sheets. Today feels like the logical endpoint of that whole series. Luna: Okay, I'm almost nervous to ask. What is it this time? Lucas: A single grape. Not a bunch, not a cluster — one individual green seedless grape, sitting on a table. Luna: You're saying a million-dollar industrial robot arm, with six degrees of freedom and sub-millimeter repeatability, can't pick up one grape? Lucas: It can try. And it will fail, or it will damage the grape, in ways that tell you almost everything about the gap between current robotics and human-level dexterity. There was a really specific study out of MIT's CSAIL lab last year — published in early 2025 — that benchmarked twelve different commercial arms on exactly this task. Luna: Wait — they actually ran a formal study on grape picking? That's both hilarious and incredibly useful. Lucas: It's a perfect test case. A grape is deformable, it's slippery, it has variable stem geometry, and it bruises if you apply more than about one newton of force — roughly the weight of an apple. The researchers wanted to isolate what I'd call the 'squeeze versus bruise' tradeoff. Luna: So the robot has to apply enough force to lift the grape, but not so much that it crushes it. That's a control problem. Lucas: Exactly. And the results were sobering. Out of twelve arms — including models from Fanuc, ABB, Kuka, and Universal Robots — only two managed to pick up a grape without visible damage more than half the time. And those two used specialized soft grippers with embedded tactile sensors. Luna: So the standard parallel-jaw gripper — which is what most factories use — just crushes the grape. Lucas: It does. But here's the interesting part: even the soft grippers failed on a specific sub-task. When the grape had a stem that was longer than five millimeters — which is common — the gripper would often grab the stem instead of the body, and the grape would swing or drop. The stem is basically a handle that the robot can't reliably distinguish from the fruit. Luna: That's a perception problem, not just a hardware problem. The robot can't tell where the stem ends and the grape begins. Lucas: Right. And that touches on the deeper issue: tactile sensing. Most industrial robots have zero tactile feedback. They have torque sensors at the joints, but not at the fingertips. So they know roughly how much force they're applying overall, but they don't know the local pressure distribution across the grape's surface. Luna: Meanwhile, a human — even a toddler — can pick up a grape without crushing it, using a combination of vision, touch, and a flexible hand that adapts in real time. Lucas: Exactly. The MIT study tried to approximate that with something called 'visuotactile sensing' — a camera embedded inside a soft gel fingertip that captures deformation patterns. When the gripper touches the grape, the camera sees how the gel deforms, and a neural network estimates the contact patch and the force. Luna: How well did that work? Lucas: It helped. The two best-performing arms both used that kind of sensor. But the success rate still topped out at around 68 percent on the first attempt. On the second attempt — after repositioning — it went up to 81 percent. But that's still a long way from reliable enough for, say, a packaging line. Luna: And in a real agricultural setting, you'd need 99.9 percent reliability. You can't have one in five grapes getting crushed or dropped. Lucas: Exactly. And that's the gap. So why is this so hard? Let's break it down into three layers. First, the physics: a grape is a thin viscoelastic shell filled with fluid. When you squeeze it, the internal pressure increases non-linearly. The relationship between gripper displacement and contact force is not linear — it changes depending on how ripe the grape is, its temperature, even its orientation. Luna: So the same gripper movement could crush a ripe grape but barely grip an underripe one. Lucas: Exactly. The second layer is control. Most industrial robots use position control — they move to a programmed coordinate. But for a deformable object, you need force control — you need to say 'apply 0.8 newtons' instead of 'move to x,y,z.' Force control is much harder to tune, and it's less common in off-the-shelf arms. Luna: And the third layer? Lucas: Perception and planning. The robot has to see the grape, estimate its shape and orientation, decide where to grasp it — ideally at the equator, not the pole — and then execute the grasp while adjusting for any slip. All in under a second if you want commercial viability. Luna: That's a lot of processing. And deep learning has made progress on some of these pieces, but not all of them. Lucas: Right. The MIT team trained a neural network on thousands of synthetic grape images and grasps. But when they tested it on real grapes, the sim to real gap was significant. The synthetic grapes were too uniform — real grapes have subtle variations in shape, translucency, and stem angle that the network hadn't seen. Luna: Classic sim to real problem. And it's worse for deformable objects because the physics simulator itself is an approximation. Lucas: Exactly. Now, you might ask: why not just use suction? A vacuum gripper could pick up a grape without squeezing it. And yes, that works — for smooth, clean, dry grapes. But in a real packing facility, grapes can be wet, dusty, or have stem remnants that break the seal. And suction grippers struggle with curved, non-porous surfaces that aren't perfectly aligned. Luna: So there's no silver bullet. Each gripper type has a failure mode. Lucas: That's the takeaway. The paper concluded that a hybrid approach — combining soft fingers with suction and tactile sensing — might get you to 95 percent reliability on table grapes. But nobody has built that yet as a commercial product. Luna: And we're not just talking about grapes. This is a proxy for a huge class of objects: berries, cherry tomatoes, mushrooms, even certain medical components like soft tissue or pills. Lucas: Exactly. The agricultural robotics market is projected to be worth around twenty billion dollars by 2030, according to a 2025 report from AgFunder. But that projection assumes we solve this kind of manipulation problem. If we can't pick a grape, we can't pick a raspberry, we can't pick a ripe peach — and that's where the money is. Luna: It's also where the labor shortage is. Farms in California and Spain are desperate for automated harvesters. But the robots still can't handle the fruit. Lucas: There is some good news. The MIT team open-sourced their dataset and their gripper designs. And since the study came out, at least three startups — I won't name them because they're still in stealth — have reached out to them about commercializing the visuotactile fingertip. Luna: So we might see a product within a couple of years. Lucas: Maybe. But the core lesson here is Humility 101 for robotics. We tend to think that if a robot can weld a car chassis or assemble a smartphone, it can do anything. But those tasks are rigid, repeatable, and engineered for the robot. A grape is none of those things. Luna: It's wild. The same arm that can place a component with micron precision can't reliably pick up a piece of fruit without mashing it. Lucas: And that's exactly the kind of conversation that keeps shows like this going — real engineering problems that don't have easy answers. If you've gotten something useful out of these episodes, a couple of dollars a month genuinely makes a difference for us. You can find that at buy me a coffee dot com slash fexingo. Luna: Yeah, it's a small thing that adds up. Keeps us ad-free and focused on the actual problems. Lucas: So — back to the grape. One of the more unexpected findings in the study was about the role of the grape's stem. The researchers found that if the stem was longer than about five millimeters, the robot's vision system often misclassified it as part of the fruit body, leading to a grasp that was off-center by a few millimeters. Luna: And a few millimeters is enough to turn a successful grasp into a failed one. Lucas: Exactly. Because the grape is only about fifteen to twenty millimeters in diameter. So a five millimeter offset is huge. The solution the team tried was a 'stem-aware' grasp planner that explicitly models the stem as a separate geometry. That improved success rates by about 12 percentage points. Luna: But that adds complexity. You need a vision system that can segment the stem from the fruit in real time, which is harder than it sounds because stems are thin and often partially occluded. Lucas: Right. And that's where things get really interesting. The team used a combination of rgb d camera and a lightweight convolutional neural network trained on a dataset of 10,000 labeled grape images. But even then, the segmentation accuracy on stems was only about 85 percent. So one in six stems was misidentified. Luna: That's a hard problem. Stems are basically thin, curved lines with low texture. They're almost designed to confuse a neural network. Lucas: And it gets harder when the grape is part of a bunch. In a cluster, stems are tangled, and the robot has to identify which stem belongs to which grape. That's a combinatorial problem that current algorithms struggle with. Luna: So the single-grape test is actually the easy case. Real-world harvesting is much messier. Lucas: Exactly. The researchers deliberately started with single grapes to isolate the grasping challenge from the perception challenge. And even that wasn't fully solved. Luna: It makes you appreciate how much of human dexterity is subconscious. We don't think about how we adjust grip force when picking up a grape — we just do it. Lucas: And that's the real frontier for robotics. Not just building a better gripper, but building a system that can sense, plan, and adapt as fluidly as a human hand. There's a reason why the most advanced robot hands today — like the Shadow Hand or the Da Vinci surgical robot — are still far less capable than a human hand in unstructured tasks. Luna: But at least now we have a benchmark. The 'grape test' could become a standard challenge in the manipulation community, like the 'coffee cup test' for mobile manipulation. Lucas: I think it already is. Since the paper came out, I've seen at least three other labs replicate the experiment with different grippers. One group at ETH Zurich used a pneumatically actuated soft gripper with embedded strain gauges and got up to 74 percent success — slightly better than the MIT result, but still not production-ready. Luna: So we're making incremental progress. But the big leap hasn't happened yet. Lucas: Not yet. And that's what makes this field so fascinating. Every time we think we've cracked manipulation, a single grape reminds us how far we have to go. Luna: I'm never going to look at a grape the same way. Lucas: Same. And maybe that's the real value of this episode — appreciating the hidden complexity in everyday objects. Next time you pop a grape into your mouth, remember: a robot can't do that.