Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Deflated Balloon
Transcript
- Lucas: So you blow up a balloon, tie it off, and hand it to a robot. The robot reaches out, closes its gripper — and the balloon squirts out sideways like a wet bar of soap. Luna: I've seen that exact thing happen in lab demos. It's almost comical how bad robots are at handling something that seems so simple. Lucas: Right. And we've spent the last hundred episodes on exactly this kind of failure — wet paper towels, loose paper, fresh eggs. But the deflated balloon is a particularly good case because it combines almost every challenge at once: it's soft, it's deformable, it's slippery, and it has no rigid structure to grab onto. Luna: Plus, if it's partially inflated, the shape changes the moment you apply force. So the robot's model of the object goes out the window the instant it touches it. Lucas: Exactly. And that's why researchers at MIT's CSAIL lab spent 2024 and early 2025 working on this exact problem. They published a paper in May 2025 — 'Deformable Object Grasping with Learned Dynamics' — where they achieved a 78 percent success rate on grasping deflated party balloons. Luna: Seventy-eight percent. So one in five still gets away. That's actually pretty good compared to what we've seen in earlier episodes. Lucas: It is. But here's the thing — they used a custom gripper with three fingers, each covered in a high-friction silicone skin, and they trained a reinforcement learning model in simulation for the equivalent of six months of real-world time. That's a lot of compute for a party trick. Luna: Why balloons though? Is there a practical application, or is this just a benchmark for deformable object manipulation? Lucas: Both. Balloons are a great proxy for things like surgical tissue, food items, or even certain types of packaging. If you can reliably grasp a deflated balloon, you can probably handle a raw chicken breast or a collapsed plastic container. Luna: So the food processing industry is a natural customer. And maybe healthcare — think about handling organs or soft tissues during surgery. Lucas: Right. But what makes the balloon so hard? Let's break it down. First, there's the issue of contact geometry. When a rigid gripper closes on a rigid object, the contact points are predictable. You can model them with high accuracy. With a balloon, the surface deforms locally, so the contact patch changes size and shape as you apply force. Luna: And the friction changes too. The balloon's latex has a low coefficient of friction, especially if there's any dust or moisture. Lucas: Exactly. The MIT team had to simulate that at a very fine scale. Their model incorporated viscoelastic material properties — that's the way the balloon's material both stretches and dissipates energy when deformed. They also modeled the air pressure inside the balloon, even for a deflated one, because there's always a little residual air that shifts during grasping. Luna: So they're essentially solving a physics problem every time the robot moves. That's computationally expensive. Lucas: It is. But they found that if they trained in simulation with randomized variations — different balloon sizes, different inflation levels, different surface textures — the policy generalized to real balloons pretty well. They went from about 40 percent success with a naive model to 78 percent with the learned dynamics model. Luna: Forty to seventy-eight is a big jump. What was the biggest source of failure in the remaining twenty-two percent? Lucas: Slippage due to unexpected deformation. In about half of the failures, the balloon would squirt out because the gripper applied too much force and deformed the balloon into a shape that provided less friction. The other half were failures where the balloon folded over itself and the gripper ended up pinching a fold instead of a stable grasp. Luna: So even with all that simulation, the robot still doesn't have a good internal model of how the balloon will behave under pressure. Lucas: Not yet. And that's the fundamental research challenge — building a model that can predict deformable object behavior in real time, with limited sensory feedback. Luna: What sensors were they using? Standard force-torque sensors? Cameras? Lucas: They used a combination — an overhead depth camera to get the initial shape of the balloon, plus tactile sensors on each fingertip that measure pressure distribution. The tactile sensors are the key. Without them, the robot has no way of knowing if it's actually making contact or if the balloon is sliding. Luna: And those tactile sensors are still relatively primitive compared to human skin. They have limited spatial resolution and they don't detect shear forces well. Lucas: Right. Human fingertips have about 2,500 mechanoreceptors per square centimeter. A typical tactile sensor grid might have 100 taxels — tactile pixels — over the same area. So the robot is essentially grasping with numb fingers. Luna: That puts the 78 percent success rate in perspective. They're doing pretty well with very limited information. Lucas: Yeah. And the MIT team is already working on the next generation — a gripper with higher-resolution tactile sensors and a learned model that updates its prediction of the balloon's shape during the grasp, not just before it. Luna: So the robot adapts in real time. That sounds like a big step. Lucas: It is. And that's the direction the whole field is moving — away from open-loop grasping, where you plan the motion and hope it works, toward closed-loop grasping where the robot continuously adjusts based on touch. Luna: I want to zoom out for a second. We've done a hundred episodes on what robot hands can't do. Do you see a timeline for when these problems get solved? Lucas: I think the next five years will see dramatic improvements in deformable object manipulation, driven by better simulations and cheaper tactile sensors. But I don't think we'll have a robot that can handle a deflated balloon as reliably as a human for at least ten years. Luna: That's sobering. But also realistic. Human dexterity is incredibly hard to replicate. Lucas: It is. And that's why we keep coming back to these seemingly trivial tasks — they're actually deep research problems that reveal the limits of our understanding. Luna: Quick honest thing — we're able to do these deep dives because a small group of listeners chips in monthly through Buy Me a Coffee dot com slash Fexingo. That's literally what funds the research time and the hosting costs for a show like this. Lucas: Yeah, it's a listener-funded model. No ads, no sponsors. Just people who find the conversations useful and want to keep them going. If that's you, we appreciate it. Luna: And if not, no pressure. The content stays free either way. Now, back to the balloon problem — there's another approach I find really interesting. Lucas: What's that? Luna: Instead of trying to grasp the balloon with fingers, some labs are working with suction grippers. A suction cup can conform to the surface and create a seal, even on soft, deformable objects. Lucas: That's a good point. Suction grippers have their own limitations — they struggle with porous surfaces, and they can't handle objects that are too curved or too flat. But for a deflated balloon, a suction cup actually works pretty well if you can get a good seal. Luna: There was a 2024 paper from the University of Tokyo that used a multi-chambered soft suction gripper. It achieved 92 percent success on a variety of deformable objects, including balloons. Lucas: I remember that one. The gripper had sixteen independently controlled chambers that could adjust the suction pattern in real time. If one chamber lost seal, the others compensated. That's a clever design. Luna: So maybe the answer isn't better fingers, but better tools. Think about how humans handle a deflated balloon — we don't just pinch it, we often trap it against our palm or use a sweeping motion. Lucas: That's a great observation. The MIT approach was essentially trying to mimic a precision pinch, which is a very human-centric strategy. But robots don't have to mimic humans. They can use different geometries and different physics. Luna: Right. And that's where the field gets really interesting — moving beyond biomimicry to robot-specific solutions. Lucas: So where does that leave us? The deflated balloon problem isn't solved, but we're seeing two clear paths forward: better sensing and modeling on one hand, and novel gripper designs on the other. Luna: And both paths are converging. The MIT team is already incorporating suction capabilities into their next gripper prototype. Lucas: I think within three years we'll see a commercially available gripper that can handle a deflated balloon with 95 percent reliability. That would unlock a lot of applications in food handling and logistics. Luna: And that's a win for everyone — except maybe the people who enjoy watching robot balloon fails on YouTube. Lucas: Ha. There's always a trade-off. But I'd rather see robots working in warehouses and kitchens than going viral for dropping things. Luna: Agreed. Next episode, we'll look at another object that gives robots fits — and maybe one that's about to be cracked. Lucas: Sounds good. See you then.