Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Deflated Balloon
Transcript
- Lucas: You ever see a child with a deflated balloon? They pick it up, stretch it, twist it, tie it — all without thinking. Luna: And a robot sees a floppy piece of rubber and basically has a meltdown. Lucas: Exactly. And that is the topic of today's episode. Why is it that in 2026, with all the progress in autonomous systems and industrial robots, a deflated balloon still stumps most robotic grippers? Luna: I mean, we've covered a lot of 'why can't robots do X' episodes — raw eggs, fitted sheets, wet paper towels. But a balloon feels almost... simpler. It's just one material. Lucas: Right, it's not like sorting mixed trash where you have infinite variability. A deflated balloon is a single object with a consistent material — latex. But it's non-rigid, it's thin, and it's highly deformable. And that combination reveals a huge gap in current robotics. Luna: So what exactly happens when a robot tries to pick up a deflated balloon? Lucas: Well, most industrial grippers are either parallel-jaw — think two fingers that close like a clamp — or suction-based. For a deflated balloon, the parallel-jaw gripper will often just squish it and the balloon slides out. The contact area is tiny and the friction isn't enough. Suction grippers, on the other hand, can't form a seal because the balloon is porous on a microscopic level and the surface isn't flat. Luna: So the robot can't even get a grip — literally. Lucas: Right. But it gets deeper. Even with more advanced grippers — like the soft robotic hands from MIT's CSAIL lab or the anthropomorphic hands from companies like Shadow Robot — the fundamental problem is sensing. A deflated balloon has no fixed shape. When you touch it, it deforms. The robot's tactile sensors have to pick up not just contact, but the changing geometry of the object in real time. Luna: And most tactile sensors, like the popular GelSight sensor, give you high-resolution surface data, but they have a limited field of view. So if the balloon shifts even a little, the contact patch moves out of the sensor's range. Lucas: Exactly. There's a 2024 paper from Berkeley that tried to use a combination of GelSight and a force-torque sensor at the wrist. They had to train a neural network to predict the balloon's deformation from partial tactile data. The success rate for picking up a deflated balloon from a flat table? About 40 percent. Luna: That's surprisingly low. I would have guessed maybe 70 percent with modern reinforcement learning. Lucas: You'd think. But the balloon's behavior is chaotic. A slight change in where the fingers touch — a few millimeters — changes the way the balloon buckles. And unlike a rigid object, the balloon doesn't have a stable grasp configuration. Every grasp is essentially unique. Luna: So is there any robot that can reliably handle a deflated balloon? Lucas: The closest I've seen is a project from the Robotic Materials group at MIT in 2025. They built a gripper with four independent fingers, each with 12 tactile sensor pixels. They also used a predictive model that simulates the balloon's deformation in milliseconds. With that setup, they achieved about 75 percent success on dry, intact latex balloons. Luna: But what if the balloon is slightly sticky or has a small hole? Real-world conditions. Lucas: Then the success rate drops to under 30 percent. The sensor data becomes noisy, and the model can't generalize. That's the rub — current systems are overfitted to lab conditions. Luna: It really highlights how far we are from human-level dexterity. A child doesn't need a model. They just feel the balloon and adjust. Lucas: And that's the deeper point. The deflated balloon is a benchmark for 'soft object manipulation' in robotics. It's not just about balloons — it's about handling anything that deforms unpredictably: a surgical suture, a piece of fabric, a food item like a tortilla. Even in warehouse automation, think about handling plastic bags or protective wraps. Luna: So solving the balloon problem could have huge spillover effects. Lucas: Absolutely. And that's why groups like the one at Toyota Research Institute are investing heavily in tactile sensing and real-time simulation. They've even started using what they call a 'balloon manipulation test' as a standard evaluation for new gripper designs. Luna: Let's talk about the simulation side. How do you even simulate a deflated balloon accurately? Lucas: It's hard. The latex is a hyperelastic material — meaning its stress-strain relationship is nonlinear. Most physics simulators in robotics, like MuJoCo or PyBullet, use linear elasticity by default. So researchers have to write custom finite element solvers that run in the loop. There's a 2025 paper from Stanford that combined a finite element solver with a learned residual model to predict balloon behavior. The simulation ran at about 50x real time — too slow for online control, but useful for training. Luna: So you train a policy in simulation, then transfer to the real robot. That's sim to real, which is common now. Lucas: Right. But the sim to real gap for deformable objects is massive. The simulation never perfectly captures the material's actual wrinkle dynamics or the surface's micro-texture. So even a great simulation-trained policy struggles when it hits a real balloon. Luna: And I imagine the balloon material itself varies — different brands, different thicknesses, different levels of moisture or static. Lucas: Exactly. That's why some labs are moving away from pure simulation and toward 'learning from demonstration' — having humans teleoperate a robot to pick up balloons, then using imitation learning. But that requires a lot of human data, and every balloon is different. Luna: It makes you wonder: is this even the right approach? Maybe we need completely new gripper designs inspired by biology. Lucas: That's a great point. A human hand has 27 degrees of freedom, thousands of mechanoreceptors, and a brain that's been fine-tuning its grasping neural networks since birth. We're not going to replicate that with a four-finger gripper and a few dozen sensors. But we don't need to. We just need to solve the specific task. Luna: So maybe the answer isn't a more human-like hand, but a task-specific tool. Like a balloon-specific gripper that uses suction and rolling. Lucas: There's actually a startup — I think it's called FlexGrip, based out of Carnegie Mellon — that's developing a gripper with a rotating roller covered in micro-suction cups. It rolls over the balloon, picks it up by creating multiple small suction points, and then rolls to release. They claim 90 percent success on latex balloons in their demos. Luna: That's promising. But is it generalizable to other soft objects? Lucas: No, and that's the trade-off. Specialization gives you performance on one task but fails on others. The holy grail is a general-purpose soft object manipulator, but we're probably a decade away. Luna: It's amazing how something as simple as a party decoration exposes the limits of robotics. Lucas: It really does. And that's why I find this topic so compelling. Every time we think robots are about to take over, we find a deflated balloon that brings them back to earth. Luna: Before we move on — Lucas, I've been thinking. This show is ad-free, and we keep it that way because of listener support. If today's conversation gave you something useful, a couple of dollars a month is genuinely what keeps these going — buy me a coffee dot com slash fexingo. Lucas: Yeah, it's a small thing that makes a big difference for us. We don't have a big network behind us, just people who find value in these deep dives. So thanks to anyone who's chipped in. Luna: Alright, back to balloons. Let's talk about what happens when you add a string. Lucas: Oh, the string is a whole other nightmare. The string is a thin, flexible, tangle-prone object that adds a new dimension of complexity. Now the robot has to deal with both the balloon's body and the string's dynamics. Luna: And if the string is knotted? We've already covered why robots can't untie knots. Lucas: Exactly. So a balloon with a string is like the final boss of soft object manipulation. But some labs are actually tackling that. There's a project at the University of Washington where they're using two robots — one holds the balloon, the other manipulates the string. They use a combination of visual tracking and force control to avoid tangling. Luna: That sounds like it requires a lot of coordination. Lucas: It does. And it's still very slow — about 30 seconds per grasp. Compare that to a human who does it in under a second. But it's a proof of concept that multi-robot systems can handle deformable objects in tandem. Luna: So where do you see the next big breakthrough coming from? Lucas: I think it's going to be from better tactile sensing combined with real-time simulation that's fast enough for closed-loop control. There's work on event-based tactile sensors that only report changes, which cuts data bandwidth. And there's progress on using neural radiance fields — NeRFs — to model object shape from partial views. If those two converge, we might see a robot that can handle a deflated balloon in the wild within five years. Luna: Five years. That's a bold prediction for something that seems so trivial to humans. Lucas: It is. But the field is moving fast. And every time we solve a problem like this, we unlock new capabilities in areas like surgery, manufacturing, and even space exploration — where you have to handle flexible materials in zero gravity. Luna: Speaking of space, how does zero gravity affect the balloon problem? Lucas: It's interesting. In microgravity, the balloon doesn't sag, so it's actually easier to predict its shape — it's just a floating film. But you lose the ability to use gravity to help you. On Earth, you can pick up a balloon by pushing down on it, creating a crease that you can pinch. In space, you have to rely entirely on friction and suction. Luna: So there's a trade-off. Easier modeling, harder grasping. Lucas: Exactly. NASA has actually funded some research on this for future robotic missions that might need to handle inflatable habitats or deployable structures. Luna: Alright, final question: if you had to design a robot today specifically to pick up a deflated balloon from a table, what would it look like? Lucas: I'd use a soft gripper with four fingers, each covered in a micro-suction material, and I'd mount a small air jet to blow the balloon slightly so it's not flat against the table. That would create a gap for the fingers to get underneath. Then I'd use a vision system to detect the balloon's boundary and plan a grasp at the edge where the material is thicker — the neck region. That would probably get you to 95 percent success. Luna: But then it wouldn't work for a balloon without a neck, like a round party balloon. Lucas: Right. So you'd need a different strategy. Maybe a combination of suction and rolling, like the FlexGrip approach. The point is, there's no one-size-fits-all solution yet. And that's what makes this such a rich research area. Luna: And a great topic for episode 95. Thanks, Lucas. Lucas: Thanks, Luna. And thanks to our listeners. If you want to see a video of a robot failing to pick up a balloon, we'll link to the Berkeley paper in the show notes. It's both painful and fascinating.