Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Raw Egg
Transcript
- Lucas: So, Luna, I want you to imagine a task that every line cook in a diner does hundreds of times a shift without thinking. One hand holds an egg, the other hand cracks it against the rim of a bowl, thumbs open the shell, and the contents drop in — no shell fragments, no burst yolk. Luna: And I'm guessing you're about to tell me that no robot arm on earth can reliably do this. Lucas: Correct. And the gap between human performance and robot performance on this one task tells you almost everything you need to know about where robotics still falls short. It's episode seventy-five of our running series on things robot arms still can't do, and today: the humble, fragile, raw chicken egg. Luna: I love this because it sounds trivial. But I know it's not. What makes it so hard? Lucas: Let me break it down into three distinct problems. First, the shell is brittle — but only along certain axes. Apply force perpendicular to the long axis, it cracks cleanly. Apply force slightly off-angle, and you either get a powder fracture or you crush the whole thing. Humans have proprioceptive feedback — we feel the shell start to give and modulate pressure in real time. A robot arm running a script doesn't 'feel' anything unless it has force-torque sensors at the wrist, and even then the control loop is running at maybe one kilohertz — fast, but not fast enough for the milliseconds it takes for a crack to propagate. Luna: So the robot is essentially blind to the moment of fracture, and by the time it registers the force drop, it's already crushed the shell? Lucas: Exactly. And that's problem one — force control bandwidth. Problem two is variation. No two eggs are identical. Shell thickness varies by breed of chicken, by age of the egg, by storage conditions. A robot that's programmed to apply exactly three newtons of force on a Tuesday morning might shatter a thin-shelled egg on Wednesday. The robot has no model for 'this is a fragile egg' unless you explicitly build one, which is why researchers at MIT's Robotic Grasping lab spent years just building a dataset of egg-cracking trajectories. Luna: I remember reading about that. They used a force-torque sensor and a high-speed camera to record thousands of human egg-cracking motions, then tried to train a neural network to reproduce them. Lucas: Right. And the network could generalize — but only to eggs of similar mass and shell thickness. The moment you handed it a jumbo egg versus a medium egg, failure rates jumped from around fifteen percent to nearly sixty percent. That's not even accounting for brown eggs versus white eggs, which have slightly different shell densities due to breed genetics. Luna: So the robot has no intuitive sense of 'this egg feels different.' It's pattern-matching on a limited dataset. Lucas: Exactly. And problem three is the most interesting one mechanically: the act of separating the shell halves after cracking. Humans use the thumbs as opposing pincers — we apply a controlled spreading force. A two-fingered gripper can do that, but the force profile is brutal. Most grippers are designed for rigid objects — they close with a preset torque. You ask a gripper to spread apart a cracked shell, it either doesn't apply enough force to separate the halves, or it slips and drops the egg. Luna: I've seen videos of robot arms trying this. The shell just kind of sticks to the gripper, or the egg slides out whole. Lucas: There's a famous clip from a 2022 IEEE conference where a robot arm spends forty seconds trying to crack an egg, and the yolk ends up on the table. But the real kicker is that even when the robot succeeds, the shell fragments are often left in the bowl. The task isn't just cracking — it's clean separation. And that requires tactile sensing at the fingertips, which most industrial grippers simply don't have. Luna: Is there any research that's actually making progress on this? Lucas: Yes. A team at UC Berkeley published a paper last year — 2025 — on visuotactile sensing. They embedded a gel-based tactile sensor into a gripper's fingertips, paired with a high-speed camera that captures deformation at two thousand frames per second. The gripper can detect the onset of a crack about fifteen milliseconds before the force sensor sees it. That's enough time to modulate the grip. Luna: Fifteen milliseconds — that's basically precognition for a robot. Lucas: It's a huge leap. But even then, their success rate on a batch of fifty eggs was seventy-two percent. Human line cooks are at ninety-nine point nine percent. So we're not there yet. And it's not just about eggs — the same sensing challenge applies to any brittle, deformable object: tomatoes, mushrooms, ceramic cups. The egg is just the perfect test case because it's cheap, consistent in shape, and brutally fragile. Luna: So what's the practical takeaway? If I'm a manufacturing engineer, should I just accept that robots can't handle fragile objects and design around it? Lucas: For now, yes. In food processing, most egg cracking is still done by specialized mechanical systems — not general-purpose robot arms. Those machines are tuned for a single egg size and orientation. But the push toward general dexterity is real. Companies like OpenAI and Google DeepMind have funded research on egg manipulation as a benchmark for tactile intelligence. The egg is becoming the new 'Rubik's cube' for robotics — a simple-seeming task that exposes every weakness in current hardware and software. Luna: Speaking of benchmarks that expose weaknesses, I've been thinking about the value of these deep dives. I mean, we spend a lot of time on what robots can't do, and that's interesting, but it's also a reminder of how much engineering ingenuity it takes to do even the simplest human tasks. Lucas: Absolutely. And it's exactly that appreciation that makes episodes like this worth creating. Honestly, if today's conversation gave you something usable — a new way to think about robotics, or just a good story to tell at dinner — that's really all we're after. And if it was worth a coffee to you, you know where to find us: buy me a coffee dot com slash fexingo. No pressure, just a way to keep this ad-free and focused on the engineering. Luna: Yeah, it's a small way to support the work. And it goes right back into research, new microphones, that kind of thing. Lucas: Exactly. So back to the egg. The Berkeley team's next step is to integrate their visuotactile sensor with a soft gripper that can conform to the egg's curvature. They believe that combination could push success rates above ninety percent within two years. But even then, the robot will still fail on the occasional thin-shelled egg — because the shell itself is a biological material, not an engineered part. Luna: And that's the underlying truth, isn't it? Biology is messy. Robots are precise. The gap is in handling mess. Lucas: Right. And that's why we'll keep coming back to these tiny, infuriating tasks. Each one teaches us something about what's missing. Next episode, we're looking at why robot arms still can't handle a single grape without bruising it. Spoiler: it's the same problem, different fruit. Luna: I'll bring the grapes. You bring the robot arm. Lucas: Deal.