Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Can't Handle an Overripe Tomato
Transcript
- Lucas: So we have done over a hundred episodes on what robot arms still can't do. And if you've been listening, you know the pattern: it's always the squishy, fragile, irregular stuff that trips them up. Luna: Right. Feathers, eggs, balloons, onions. We've covered a lot of produce. Lucas: Today I want to talk about the tomato. Specifically, an overripe tomato. The kind that's maybe a day past its prime—soft, a little give in the skin, but still whole. Luna: The one you'd be careful not to squeeze too hard at the grocery store. Lucas: Exactly. And the reason I'm picking this is a 2025 study from MIT's CSAIL lab that benchmarked twelve commercial grippers on tomatoes at different ripeness levels. The best success rate they got was forty-seven percent. Luna: So less than half the time they didn't crush it or drop it? Lucas: Correct. And that's with the grippers tuned specifically for that task. Not a general-purpose pick and place. They spent hours calibrating force thresholds. Luna: What's the fundamental physics problem here? Because it seems like a simple thing—just grip gently enough. Lucas: The problem is that a tomato is deforming as soon as you touch it. So the robot's force sensors see a changing geometry. The gripper either applies too much pressure and squishes it, or too little and loses grip. And the really insidious thing is that the optimal force window is tiny—maybe a few tenths of a newton wide for an overripe one. Luna: So it's not just about sensor resolution. It's about real-time adaptation. Lucas: Right. Human hands have about seventeen thousand touch receptors in the palm and fingers. A typical robotic tactile sensor has maybe a few hundred taxels. And our brain processes that data with millions of years of evolution. The robot has a control loop running at maybe a kilohertz. Luna: What about machine learning? Can't a neural net learn the right pressure profile? Lucas: People have tried. The problem is that the training data is hard to generate. You can't just collect a million successful tomato grasps because you'll destroy a million tomatoes. And the variance is huge—every tomato has a slightly different shape, stiffness, surface texture. Luna: So simulators? You could simulate a deformable tomato. Lucas: Simulation of soft bodies is notoriously compute-intensive and still doesn't match reality. The MIT team tried sim to real transfer and saw a thirty percent drop in success rate compared to real-world calibration. Luna: Ouch. So what's the most promising approach you've seen? Lucas: There's a group in Japan—at the University of Tokyo—that built a gripper using water-filled silicone bags. The bags conform to the object shape passively, like a water balloon. The robot only needs to apply a very gentle overall force because the water distributes pressure evenly. Luna: So it's a mechanical solution rather than a control one. Lucas: Exactly. They reported eighty-two percent success on overripe tomatoes. But the gripper is slow—it takes about two seconds to conform, which is an eternity on a production line. Luna: Two seconds per tomato. At that rate, you'd need a lot of parallel arms to keep up with a packing plant. Lucas: Right, and the water-filled bags are also prone to leaking after a few thousand cycles. So the maintenance cost goes up. Luna: This feels like one of those problems where hardware innovation might outpace software. The sensor arrays just aren't dense enough yet. Lucas: Some researchers are working on neuromorphic tactile sensors that mimic the human touch receptor density. But those are still lab prototypes. Commercially, we're probably five to ten years away from a tomato that a robot can handle as reliably as a human. Luna: And that matters beyond tomatoes. It's the same problem for any soft fruit, for certain medical devices, for handling biological tissue in labs. Lucas: Absolutely. The tomato is a stand-in for a whole class of objects. If you can solve the tomato, you can solve a lot of the manipulation economy. Luna: It's interesting that after a hundred and fourteen episodes, this is still frontier research. Lucas: That's partly why we keep doing these episodes. There's no shortage of things robot arms still can't do. And if today's conversation gave you something usable—a new way to think about the challenge—that's the whole point. We deliberately don't run ads on these shows. If you want to support that choice, the link is buy me a coffee dot com slash fexingo. Luna: Yeah, it's a clean way to keep the content focused on the curiosity rather than the commercial break. Lucas: Alright, back to the tomato. One more angle I found compelling: the MIT study also looked at how humans perform the same task. They had participants pick up tomatoes of varying ripeness with a standard industrial gripper—but with haptic feedback through a controller. Luna: So teleoperation, essentially. Lucas: Yes. And humans with just a few minutes of practice achieved ninety-four percent success. That suggests the control problem isn't just about hardware—it's about the richness of feedback. Luna: Because we can feel the tomato starting to give and adjust instantly. Lucas: Right. And the robot doesn't have that proprioceptive loop. It measures force at discrete points, but it can't feel the gradual change in compliance across the whole surface. Luna: So one path forward is making teleoperation easier, then using that human data to train autonomous systems. Lucas: That's exactly what a startup called SoftGrip is doing. They have a glove with embedded sensors that maps human hand forces to a robotic gripper. They've collected over ten thousand human tomato grasps and used that to train a reinforcement learning model. Luna: What's their success rate now? Lucas: They claim seventy-one percent in the lab. Not yet commercially viable, but it's a step change from the forty-seven percent the benchmark achieved. Luna: So we're moving, but slowly. Lucas: Slowly is the right word. But every percent improvement in grasping success has a massive economic impact. A one percent reduction in bruised produce saves the global food industry an estimated two hundred million dollars a year. Luna: That's the kind of number that keeps funding coming into this research. Lucas: For sure. And that's why I think we'll see a commercial tomato-handling robot within the decade. The economic incentive is just too strong. Luna: I'm curious to see whether it ends up being a sensor breakthrough or a mechanical one. Lucas: My bet is on a hybrid. Better sensors to read the tomato, and a compliant gripper that doesn't fight it. But we'll have to do another episode in five years and see. Luna: I look forward to that. In the meantime, I'll be careful with my tomatoes. Lucas: And I'll be watching the CSAIL preprint server.