Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Coffee Cup
Transcript
- Lucas: So here's a thing you probably do without thinking, maybe three or four times a day: you pick up a coffee cup. Ceramic mug, maybe a handle, maybe not, hot liquid inside. Your brain and hand coordinate in milliseconds. A robot arm, even a million-dollar industrial one, still can't do it reliably. Luna: I've definitely seen those viral videos of robot arms smashing mugs or just pushing them across the table. But what exactly makes a coffee cup so hard? Lucas: Great question. Let's start with the handle. Most robot grippers are designed for parallel jaw grasping — two flat or padded surfaces squeezing an object. A mug handle is a thin curved loop. To grasp it with a parallel gripper, you'd have to align perfectly inside that loop, which is maybe two centimeters wide. Miss by a millimeter and the gripper either bumps the mug body or just closes on air. Luna: Right, and the handle's orientation varies. Some mugs have a D-shaped handle, others a C-shape, some have no handle at all. Lucas: Exactly. So the first challenge is geometric variability. But even if you solve that — say you use a three-fingered hand — you now face the material problem. Ceramic is hard, smooth, and glazed. The friction coefficient between a typical silicone gripper pad and a glazed ceramic surface is surprisingly low. Squeeze too gently and the mug slips. Squeeze too hard and you risk cracking the ceramic or crushing a thin-walled mug. Luna: And that's before you add heat. A fresh cup of coffee can be 70 degrees Celsius. Hot enough to degrade many soft gripper materials over time or cause thermal expansion that changes the grip geometry. Lucas: Right. A 2025 paper from MIT's soft robotics lab specifically looked at heat-resistant grippers for hot beverage handling. They found that standard silicone pads lose about 40 percent of their gripping force when the mug surface reaches 60 degrees Celsius. The material softens, the friction drops, and the mug starts to rotate in the grasp. Luna: So the gripper needs to be simultaneously compliant enough to conform to the handle shape, yet stiff enough to maintain grip when hot. That's a material science trade-off. Lucas: And it gets worse. Consider the liquid inside. When you pick up a full mug, the coffee sloshes. That shifting center of mass changes the torque the gripper has to counteract. A robot arm that's programmed to lift with a constant force profile will see the mug suddenly tilt as the liquid moves. Without real-time force feedback, it either drops the mug or overcompensates and spills. Luna: That's why most robot barista projects, like the ones from Cafe X or Briggo, use a fixed cup holder and just move the espresso machine's spout. They never actually have the robot pick up a mug. Lucas: Exactly. They sidestep the whole manipulation problem. But if we want a robot to clear a table or bring you a coffee from the kitchen, it has to pick up the mug. And that means solving what roboticists call 'form closure' versus 'force closure.' Form closure means the gripper geometry cradles the object so it can't move even without friction — like a bucket handle in a hook. Force closure relies on friction and squeeze force. A mug handle is small and curved, so achieving form closure with a typical gripper is almost impossible. You're stuck with force closure, which is fragile. Luna: There's been some interesting work with underactuated hands — hands that use fewer motors than joints, so they passively conform to shapes. The Yale OpenHand project has a model that can wrap around a mug body, not just the handle. But even that struggles with the handle's geometry. Lucas: Right. And that approach works better for mug bodies — you can envelope the cylinder. But then you have to deal with the handle as an obstacle. If you grip the body, the handle might bump into the gripper. So you need to plan the grasp around the handle's location, which requires vision and a model of the mug. Most commercial systems don't have that level of perception. Lucas: Now, let's talk about what happens when you actually try to lift a mug with a handle. A team at the University of Tokyo in 2023 published a system where a robot arm used a combination of vision and tactile sensing to locate the handle, approach from the correct angle, and insert a gripper finger into the loop. They achieved an 80 percent success rate on the same mug every time. But when they switched to a different mug — different handle shape, different weight — the success rate dropped to 30 percent. Luna: So it's not robust to variance. That's a huge problem for real-world use. You can't tell your robot, 'only pick up this one specific mug.' Lucas: Precisely. And that's the deeper point: the coffee cup problem is a microcosm of the entire manipulation challenge in robotics. It combines geometry uncertainty, material properties, thermal effects, dynamic loads, and the need for precision. Get all those right, and you've basically solved a massive chunk of general object manipulation. Lucas: But here's the thing — if you found today's deep dive useful, this is exactly the kind of ad-free, detailed conversation that listener support makes possible. We keep this show free of sponsors and commercials, and the way that works is people like you who get value from it toss a few bucks our way. You can do that at buy me a coffee dot com slash fexingo. That's buy me a coffee dot com slash fexingo. No pressure, just if it's worth a coffee to you. Luna: Yeah, and honestly, we love diving into these specific problems that don't get covered in the headlines. Knowing that listeners appreciate that keeps us going. Lucas: Alright, back to the mug. So we've got geometry, friction, heat, and sloshing. What about the act of actually setting the mug down? That's surprisingly hard too. You have to decelerate precisely, make contact without jarring the mug, and release the grip at exactly the right moment. Too early, the mug tips. Too late, you're pushing it into the table. Luna: And if you're placing it on a coaster or a saucer, you need millimeter accuracy. Most industrial robots have repeatability of about 0.1 millimeters, but that's in a controlled environment with a known object. With a mug you just picked up, the arm's kinematics might shift slightly due to the load. Lucas: Right. And there's also the issue of the mug's base. Many mugs have a slight taper — wider at the bottom than the top. So if you grip the body and then set it down, the mug might rock if the gripper doesn't release cleanly. Some researchers have proposed using a flexible wrist that can passively align during placement, but that adds complexity. Luna: I remember a paper from the IEEE International Conference on Robotics and Automation last year where they used a suction cup on the mug's bottom to lift it from above. That avoids the handle entirely. But then you need a flat, clean surface on the bottom of the mug, which isn't always the case. Lucas: And suction fails with textured or wet bottoms. So that's not universal either. The reality is, we don't have a single gripper that can handle all mugs in all conditions. That's why you don't see robot waiters in real restaurants yet — despite all the hype. The closest we have are systems like the ones from Bear Robotics that carry trays with fixed cup positions, but they don't pick up individual mugs. Luna: So what's the most promising approach right now? I've seen some work on using granular jamming grippers — bags of coffee grounds that stiffen when air is sucked out. They can conform to any shape, including a mug handle. Lucas: That's a fascinating direction. The gripper from the University of Chicago's lab can envelop a mug handle and then jam to lock into shape. But the jamming process takes a few seconds, and the grip strength is limited. For heavy mugs, it might not be enough. Also, the coffee grounds can leave residue on the mug. Lucas: Another approach is using adhesive grippers — gecko-inspired dry adhesives that stick to smooth surfaces without residue. They work on ceramic, but they lose adhesion with dust or moisture. And they don't handle the torque from a sloshing liquid well. Luna: So it seems like we're still in the research phase. Is there any commercial product that can reliably pick up an arbitrary coffee cup? Lucas: Not that I know of. The closest might be the Shadow Robot Dexterous Hand, which has 24 joints and can perform a precision grasp on a mug handle. But it costs over $100,000 per hand, requires extensive programming, and still fails on unfamiliar mugs. So it's not practical for widespread use. Luna: That's a stark reminder of how far we are from general-purpose manipulation. A task a four-year-old can do reliably is still a frontier research problem. Lucas: Exactly. And that's why this is such a rich area. Every failure teaches us something about perception, materials, control, and design. I expect we'll see breakthroughs in the next five years as tactile sensors become cheaper and machine learning models get better at generalizing across objects. Luna: I'm curious — if you had to bet on one technology that gets us there first, what would it be? Lucas: I'd put my money on soft, sensor-rich hands combined with learning from demonstration. Train a neural network on thousands of mug-grasping attempts in simulation, then transfer to a real robot with tactile feedback. The simulation to reality gap is closing fast. Maybe in five years, a robot will be able to pick up any mug in your kitchen with 95 percent reliability. Luna: That would be a milestone. Until then, I'll keep fetching my own coffee. Lucas: And appreciate how hard that simple action really is.