Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Handle a Balloon
Transcript
- Lucas: So, I tried something over the weekend. I blew up a balloon — not a party thing, more a 'let me watch my robot arm fail at something' thing. Luna: Okay, I need to hear this. What happened? Lucas: I set up a standard six-axis arm with a two-finger parallel gripper. The balloon was just sitting on a table, inflated, maybe 20 centimeters across. I told the arm to pick it up. Luna: Let me guess — it either couldn't grasp it, or it popped it immediately. Lucas: Both, actually. First try, the gripper closed and the balloon just shot out sideways like a watermelon seed. Second try, I increased grip force — and pop. So now I'm thinking: why is this so hard? Turns out, a balloon is a perfect storm of bad properties for a robot. Luna: What specifically makes it harder than, say, an egg? Eggs are fragile, but we've seen robots handle eggs. Lucas: Great question. An egg has a rigid shell. You can model its shape, you know exactly where to apply force, and the shell distributes pressure. A balloon is the opposite: it's a thin elastic membrane with no structural rigidity. The shape deforms the moment you touch it. The surface is almost frictionless — latex has a coefficient of friction around 0.04 against most gripper materials. And the internal pressure is low, so any local force causes large local deformation. Luna: So it's not just about force control. It's about the feedback loop breaking down. Lucas: Exactly. Most industrial grippers use force-torque sensors at the wrist. They measure the net force and moment at the tool flange. But when you're dealing with a balloon, the forces are below the sensor's noise floor — we're talking less than 0.1 Newtons. A typical FT sensor has a resolution of about 0.05 Newtons, but in practice, with vibration and thermal drift, you can't reliably measure below 0.2 Newtons. So the arm is essentially blind. Luna: I remember reading a paper from MIT CSAIL — I think it was 2024 or early 2025 — where they tried to get a robot to inflate a balloon by holding it against a nozzle. They had a similar problem. The arm couldn't tell if it was gripping too tightly or too loosely. Lucas: That's the exact paper I'm thinking of. Led by someone in Russ Tedrake's group. They used a Kuka LBR iiwa with joint torque sensors — which are better than wrist FT sensors, but still, the balloon slipped off the nozzle something like 60 percent of the time in early trials. Luna: It's a hard problem because humans do it effortlessly. You pick up a balloon without thinking. You modulate grip pressure in real time based on tactile cues — the slight give, the tension change. Lucas: Right. And that's the deeper point. For a robot to handle a balloon, you'd need tactile sensors with high spatial resolution and high sensitivity across the entire gripping surface. Not just at the wrist. You need to feel the local pressure distribution at each fingertip. And you need control loops that can update at kilohertz rates, not just the typical hundred-hertz servo loop. Luna: There's also the sheer variety of balloons. Different latex thicknesses, different inflation pressures, different shapes. A long twisting balloon is a different object from a round party balloon. A water balloon is even worse. Luna: I want to ask — have any teams actually succeeded? Or is this still an open research problem? Lucas: There's been progress. The MIT group eventually got their system to work about 80 percent of the time by using a soft pneumatic gripper with embedded capacitive sensors. The gripper conforms to the balloon's shape passively, and the sensors measure local strain. They used a model-predictive control scheme that adjusts grip force based on predicted deformation. It's clever, but it's also highly tuned to that specific task. Luna: So it doesn't generalise to, say, picking up a piece of fruit or a delicate electronic component. Lucas: Not yet. And that's the frontier. The balloon problem exposes a core limitation in current robotics: we're good at handling rigid, well-characterised objects in structured environments. We're bad at handling compliant, low-friction, deformable objects that are hard to model. Luna: And this isn't just a lab curiosity. I know DARPA ran a challenge in 2025 focused on delicate manipulation for medical applications. One task involved a robot arm moving an inflated balloon across a table without popping it, simulating handling a blood vessel or a catheter. Lucas: I saw the results of that. The best-performing team used a combination of vision-based tactile sensors — basically a camera inside a gel fingertip — and a reinforcement learning policy trained in simulation. They achieved a 70 percent success rate over 100 trials. That's impressive, but still not reliable enough for surgery. Luna: It's interesting because we often think of robotics progress in terms of big, dramatic tasks — walking, flying, manipulating heavy objects. But the real frontier might be these tiny, subtle interactions. Lucas: Speaking of subtle interactions — you know, this show exists because a number of listeners chip in monthly through buy me a coffee dot com slash fexingo. It's listener-supported, no ads, and that support is what lets us spend time on these deep dives. Luna: Yeah, and we really appreciate it. Keeps us independent and lets us pick topics that are genuinely interesting rather than just clickable. Lucas: So if you find value in these conversations, that's the place. Now, back to balloons — one thing I didn't mention is that the inflation level matters enormously. A loosely inflated balloon is easier to grip because it deforms more and distributes force better. A fully taut balloon has very little give, so the local stress at the contact point is much higher. Luna: Did you try different inflation levels in your weekend experiment? Lucas: I did. I inflated one balloon to about 80 percent of maximum diameter, and another to maybe 40 percent. The under-inflated one was actually graspable — the gripper could sink into the material a bit. But the fully inflated one? No chance. It either slipped or popped. Luna: That aligns with something I read about picking up plastic bags. Same problem — low stiffness, unpredictable deformation. Lucas: Plastic bags are another classic. There was a paper from UC Berkeley a few years ago — they used a suction gripper combined with a pinch grasp, and even then, success rate was only around 50 percent for crumpled bags. The balloon is harder because you can't use suction — the surface is curved and smooth, and a vacuum would just collapse it. Luna: So what would a truly robust balloon-handling system require? Let's say we want a robot that can pick up any balloon, any inflation level, reliably. Lucas: You'd need three things. First, high-resolution tactile sensing across the entire gripper surface — think hundreds of taxels per finger, sensitive to forces below 0.01 Newtons. Second, a gripper that can actively conform its shape — something like a granular jamming gripper that can go from soft to rigid on demand. And third, a control policy that can adapt in real time based on tactile feedback, not just preprogrammed force thresholds. Luna: None of which exists off the shelf today, right? Lucas: Not in an integrated package. The components exist in labs — there are tactile sensors with enough resolution, there are jamming grippers, there are learning-based controllers. But combining them into a reliable system that works across a broad range of objects is still a research challenge. Luna: It's almost a metaphor for the state of robotics. We can build cars, we can build assembly lines, but we can't yet build a robot that can reliably pick up a party decoration. Lucas: Exactly. And until we solve that — handling the light, the soft, the slippery, the deformable — there's a whole category of tasks that remain out of reach. Harvesting delicate fruit, handling biological tissue, sorting recyclables, folding laundry. These aren't exotic tasks; they're everyday things that humans do without thinking. Luna: And the balloon is a perfect test case because it combines all those challenges in one object. If you can handle a balloon, you can probably handle a tomato, a piece of silk, a raw egg. Lucas: That's why I think it's a great benchmark for the field. There's actually a proposal circulating in the robotics community — some researchers want to make 'the balloon test' a standard manipulation benchmark, like the peg-in-hole task was for assembly. Luna: I'd support that. It's easy to replicate, it's cheap, and it's genuinely hard. Plus, it's visually dramatic — a robot popping a balloon gets people's attention. Lucas: It does. And it highlights how far we still have to go. But also how far we've come — ten years ago, we were celebrating robots that could pick up a block. Now we're frustrated that they can't pick up a balloon. That's progress, in its own way. Luna: So what's the next step? Are you going to build your own tactile gripper? Lucas: I'm tempted. There are open-source designs for capacitive tactile sensors that you can 3D print with conductive filament. I might try to integrate one with a soft gripper and see if I can beat my 0-for-10 record. Luna: You have to film that. I want to see the moment it finally works. Lucas: If I get it to work, it'll be an episode. And if I pop a hundred balloons in the process, at least I'll have confetti. Luna: Alright, we'll hold you to that. Thanks for listening, everyone. Lucas: Yeah, until next time.