Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Can't Flip a Pancake
Transcript
- Lucas: You know what sounds simple? Flipping a pancake. You pour batter, wait for bubbles, slide a spatula under, and flick your wrist. Done. Luna: And yet for a robot arm, that flick of the wrist is a nightmare of physics. Lucas: Exactly. It's a dynamic manipulation task — the robot has to apply a controlled impulse to launch the pancake, then catch it as it falls, all while the pancake deforms and slides on the spatula. Most industrial robots are designed for rigid, repeatable motions. Pancakes are the opposite. Luna: So we're not talking about a factory floor problem. This is more like a lab curiosity, right? Lucas: It's a benchmark. If you can reliably flip a pancake, you've solved a whole class of problems: flipping a burger, folding a piece of fabric, even manipulating a flexible circuit board. The MIT CSAIL group published a paper in early 2025 where they used a soft gripper — basically a pneumatic balloon — and trained a neural network to learn the flip through trial and error. They got to a 73 percent success rate on a standard stovetop. Luna: Seventy-three percent is decent, but not good enough for a Sunday brunch. What goes wrong the other 27 percent of the time? Lucas: A few failure modes. The pancake sticks to the spatula, so it doesn't launch cleanly. Or it folds over on itself midair — the aerodynamics of a deformable disk are tricky. Or the robot misjudges the catch timing and the pancake lands half off the spatula. The MIT team used high-speed cameras running at 200 frames per second to track the pancake's trajectory, and they still saw failures. Luna: Two hundred frames per second just to catch a pancake. That's insane processing requirements. Lucas: Right. And they had to model the pancake's material properties in real time — how it stretches, how it releases from the surface. That's a lot of compute for what a human does unconsciously. The key breakthrough in the MIT work wasn't just the gripper; it was the reinforcement learning algorithm that learned from failure. They let the robot flip thousands of pancakes in simulation, then transferred to a real kitchen setup. Luna: Sim to real transfer for pancake flipping. That's impressive, but still not ready for commercial use. Lucas: Not even close. But it tells us where the research frontier is. And it also highlights why these projects cost a lot — between the custom hardware, the simulation licenses, and the compute time, you're looking at hundreds of thousands of dollars per lab per year. Luna: That's steep. But for listeners who geek out on this stuff — and I know many of you do — if today's conversation gave you something usable, there's a way to support this kind of ad-free deep dive. Lucas: Yeah, the show stays ad-free because of listener support. If you found the pancake physics interesting, or you just want to keep these episodes coming, head over to buy me a coffee dot com slash fexingo. It's a simple way to say 'keep going.' Luna: And we really appreciate it. So back to the flip: what's the current state of the art, Lucas? Lucas: Well, since the MIT paper, a team at CMU has been working on a multi-modal gripper that combines suction and soft fingers. They claim a 91 percent success rate on pancakes, but only under controlled conditions — same batter consistency, same pan temperature, same flip height. Change any variable and it drops. Luna: So robustness is the real issue. A human can adapt to a thicker batter or a hotter pan instantly. Lucas: Exactly. The robot's model is brittle. It learns the specific dynamics of one setup. If you change the spatula material — say from metal to silicone — the friction changes, and the whole flip trajectory needs retraining. That's the big open problem: generalizable manipulation. Luna: Is anyone working on that? Lucas: Yes, there's a fascinating approach from Google DeepMind called 'universal dexterous manipulation.' They're training a single policy on hundreds of household objects — eggs, pancakes, shirts, wires — using a massive simulator called DManip. But it's early. They can flip a pancake in simulation, but sim to real transfer is still a 50 percent success rate. Luna: Fifty percent is worse than MIT's 73 percent on the real robot. Lucas: That's because MIT carefully tuned their sim to match their exact hardware. DeepMind's sim is more generic, so the gap is larger. But the idea is that once they bridge that gap, the same policy could flip a pancake, fold a towel, and pick up a mushroom without re-training. Luna: That would be a game changer for manufacturing and logistics. Imagine a robot arm that can handle flexible packaging, cables, or food items without constant reprogramming. Lucas: Exactly. The pancake is just a proxy. But it's a good one because it combines so many challenges: adhesion, deformation, aerodynamics, timing, and force control. If you can master the pancake, you've got a shot at a lot of real-world tasks. Luna: What about the hardware side? Are there any new robot hands specifically designed for flipping? Lucas: A few. The Shadow Dexterous Hand has been modified with a flexible palm that can cup the pancake. And there's a startup called Moley Robotics that built a full kitchen robot with a spatula tool changer — it can switch between frying spatula, flipper, and turner. But their flipping demo still shows about one in ten pancakes falling on the floor. Luna: That's not ideal for a commercial kitchen. You'd need near-perfect reliability. Lucas: Right. And safety — a robot arm swinging a hot spatula near a human is a hazard. So they usually operate in a gated cell. That limits adoption. Luna: So where do you see the next big breakthrough? Sensing or control? Lucas: I think it's sensing. Specifically, tactile sensing that gives the robot information about slip and pressure in real time. Most current grippers have no sense of touch — they just know position and torque. If you add a dense array of capacitive or piezoelectric sensors, the robot can feel the pancake starting to slide and adjust its grip mid-flip. That's where groups like UC Berkeley are focusing. Luna: That makes sense. Humans rely on touch feedback constantly — we feel the spatula angle and adjust. Lucas: Exactly. The Berkeley team has a sensorized spatula with 16 taxels — tactile pixels — that measure shear force. They're integrating it with a learning-based controller and claim a 94 percent success rate on flipping crepes, which are even thinner and more delicate than pancakes. Luna: Crepes are a harder test because they're more fragile. So 94 percent is impressive. Lucas: It is, but they're still in the lab. The real world has variable batter, dirty pans, and greasy spatulas. So there's a long way to go before your local diner installs a robot short-order cook. Luna: And even then, would we want that? Part of the charm of a diner is the human cook flipping pancakes with style. Lucas: That's a fair point. The goal isn't to replace people; it's to automate the dangerous, repetitive tasks. Maybe the robot handles the first hundred flips during a breakfast rush, and the human does the finishing touches. That hybrid approach is where many roboticists are headed. Luna: So the pancake is a symbol of the broader challenge in robotics: making machines that can handle the messiness of the physical world. Lucas: Exactly. And every time a robot drops a pancake, we learn something about control, sensing, and materials. It might take another decade, but I think we'll get there. Until then, I'll keep flipping my own pancakes. Luna: And I'll keep enjoying them. Thanks for the great conversation, Lucas. Lucas: Thanks, Luna. And thanks to our listeners for sticking with us. That's it for this episode of The Robotics Podcast.